diff --git "a/results.json" "b/results.json"
--- "a/results.json"
+++ "b/results.json"
@@ -209,6 +209,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ar",
@@ -321,6 +328,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"as",
@@ -356,6 +370,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"awa",
@@ -391,6 +412,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"az",
@@ -426,6 +454,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"be",
@@ -461,6 +496,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"bho",
@@ -615,6 +657,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ceb",
@@ -650,6 +699,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ckb",
@@ -930,6 +986,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"fa",
@@ -1098,6 +1161,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"gu",
@@ -1182,6 +1252,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"hi",
@@ -1336,6 +1413,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"id",
@@ -1427,6 +1511,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ilo",
@@ -1560,6 +1651,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"jv",
@@ -1630,6 +1728,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"kk",
@@ -1665,6 +1770,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"km",
@@ -1700,6 +1812,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"kn",
@@ -1735,6 +1854,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ko",
@@ -1847,6 +1973,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"mai",
@@ -1882,6 +2015,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"mg",
@@ -1924,6 +2064,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ml",
@@ -1959,6 +2106,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"mr",
@@ -1994,6 +2148,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ms",
@@ -2036,6 +2197,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"my",
@@ -2071,6 +2239,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ne",
@@ -2162,6 +2337,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ny",
@@ -2253,6 +2435,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"or",
@@ -2288,6 +2477,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"pa",
@@ -2365,6 +2561,13 @@
     "metric":"chrf",
     "score":0.5129096175
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"pt",
@@ -2372,6 +2575,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"pt",
@@ -2407,6 +2617,13 @@
     "metric":"chrf",
     "score":0.632650236
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ro",
@@ -2561,6 +2778,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"sd",
@@ -2596,6 +2820,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"si",
@@ -2687,6 +2918,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"so",
@@ -2729,6 +2967,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"sr",
@@ -2771,6 +3016,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"su",
@@ -2904,6 +3156,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ta",
@@ -3065,6 +3324,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ti",
@@ -3100,6 +3366,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"tr",
@@ -3142,6 +3415,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"uk",
@@ -3219,6 +3499,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"ur",
@@ -3254,6 +3541,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"uz",
@@ -3289,6 +3583,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"vi",
@@ -3513,6 +3814,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"amazon\/nova-micro-v1",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"amazon\/nova-micro-v1",
     "bcp_47":"yue",
@@ -3849,6 +4157,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ar",
@@ -3961,6 +4276,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"as",
@@ -3996,6 +4318,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"awa",
@@ -4031,6 +4360,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"az",
@@ -4066,6 +4402,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"be",
@@ -4101,6 +4444,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"bho",
@@ -4255,6 +4605,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ceb",
@@ -4290,6 +4647,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ckb",
@@ -4570,6 +4934,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"fa",
@@ -4738,6 +5109,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"gu",
@@ -4822,6 +5200,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"hi",
@@ -4976,6 +5361,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"id",
@@ -5067,6 +5459,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ilo",
@@ -5200,6 +5599,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"jv",
@@ -5270,6 +5676,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"kk",
@@ -5305,6 +5718,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"km",
@@ -5340,6 +5760,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"kn",
@@ -5375,6 +5802,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ko",
@@ -5487,6 +5921,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"mai",
@@ -5522,6 +5963,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"mg",
@@ -5564,6 +6012,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ml",
@@ -5599,6 +6054,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"mr",
@@ -5634,6 +6096,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ms",
@@ -5676,6 +6145,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"my",
@@ -5711,6 +6187,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ne",
@@ -5802,6 +6285,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ny",
@@ -5893,6 +6383,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"or",
@@ -5928,6 +6425,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"pa",
@@ -6005,6 +6509,13 @@
     "metric":"chrf",
     "score":0.5811687089
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"pt",
@@ -6012,6 +6523,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"pt",
@@ -6047,6 +6565,13 @@
     "metric":"chrf",
     "score":0.7440224371
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ro",
@@ -6201,6 +6726,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"sd",
@@ -6236,6 +6768,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"si",
@@ -6327,6 +6866,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"so",
@@ -6369,6 +6915,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"sr",
@@ -6411,6 +6964,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"su",
@@ -6544,6 +7104,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ta",
@@ -6705,6 +7272,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ti",
@@ -6740,6 +7314,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"tr",
@@ -6782,6 +7363,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"uk",
@@ -6859,6 +7447,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"ur",
@@ -6894,6 +7489,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"uz",
@@ -6929,6 +7531,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"vi",
@@ -7153,6 +7762,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.5-sonnet",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"anthropic\/claude-3.5-sonnet",
     "bcp_47":"yue",
@@ -7489,6 +8105,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ar",
@@ -7601,6 +8224,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"as",
@@ -7636,6 +8266,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"awa",
@@ -7671,6 +8308,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"az",
@@ -7706,6 +8350,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"be",
@@ -7741,6 +8392,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"bho",
@@ -7895,6 +8553,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ceb",
@@ -7930,6 +8595,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ckb",
@@ -8210,6 +8882,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"fa",
@@ -8378,6 +9057,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"gu",
@@ -8462,6 +9148,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"hi",
@@ -8616,6 +9309,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"id",
@@ -8707,6 +9407,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ilo",
@@ -8840,6 +9547,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"jv",
@@ -8910,6 +9624,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"kk",
@@ -8945,6 +9666,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"km",
@@ -8980,6 +9708,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"kn",
@@ -9015,6 +9750,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ko",
@@ -9127,6 +9869,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"mai",
@@ -9162,6 +9911,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"mg",
@@ -9204,6 +9960,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ml",
@@ -9239,6 +10002,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"mr",
@@ -9274,6 +10044,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ms",
@@ -9316,6 +10093,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"my",
@@ -9351,6 +10135,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ne",
@@ -9442,6 +10233,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ny",
@@ -9533,6 +10331,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"or",
@@ -9568,6 +10373,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"pa",
@@ -9645,6 +10457,13 @@
     "metric":"chrf",
     "score":0.5826652331
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"pt",
@@ -9652,6 +10471,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"pt",
@@ -9687,6 +10513,13 @@
     "metric":"chrf",
     "score":0.7292524118
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ro",
@@ -9841,6 +10674,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"sd",
@@ -9876,6 +10716,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"si",
@@ -9967,6 +10814,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"so",
@@ -10009,6 +10863,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"sr",
@@ -10051,6 +10912,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"su",
@@ -10184,6 +11052,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ta",
@@ -10345,6 +11220,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ti",
@@ -10380,6 +11262,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"tr",
@@ -10422,6 +11311,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"uk",
@@ -10499,6 +11395,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"ur",
@@ -10534,6 +11437,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"uz",
@@ -10569,6 +11479,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"vi",
@@ -10793,6 +11710,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-3.7-sonnet",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"anthropic\/claude-3.7-sonnet",
     "bcp_47":"yue",
@@ -11129,6 +12053,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ar",
@@ -11241,6 +12172,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"as",
@@ -11276,6 +12214,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"awa",
@@ -11311,6 +12256,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"az",
@@ -11346,6 +12298,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"be",
@@ -11381,6 +12340,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"bho",
@@ -11535,6 +12501,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ceb",
@@ -11570,6 +12543,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ckb",
@@ -11850,6 +12830,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"fa",
@@ -12018,6 +13005,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"gu",
@@ -12102,6 +13096,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"hi",
@@ -12256,6 +13257,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"id",
@@ -12347,6 +13355,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ilo",
@@ -12480,6 +13495,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"jv",
@@ -12550,6 +13572,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"kk",
@@ -12585,6 +13614,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"km",
@@ -12620,6 +13656,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"kn",
@@ -12655,6 +13698,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ko",
@@ -12767,6 +13817,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"mai",
@@ -12802,6 +13859,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"mg",
@@ -12844,6 +13908,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ml",
@@ -12879,6 +13950,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"mr",
@@ -12914,6 +13992,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ms",
@@ -12956,6 +14041,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"my",
@@ -12991,6 +14083,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ne",
@@ -13082,6 +14181,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ny",
@@ -13173,6 +14279,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"or",
@@ -13208,6 +14321,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"pa",
@@ -13285,6 +14405,13 @@
     "metric":"chrf",
     "score":0.5693090483
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"pt",
@@ -13292,6 +14419,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"pt",
@@ -13327,6 +14461,13 @@
     "metric":"chrf",
     "score":0.6981449573
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ro",
@@ -13481,6 +14622,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"sd",
@@ -13516,6 +14664,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"si",
@@ -13607,6 +14762,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"so",
@@ -13649,6 +14811,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"sr",
@@ -13691,6 +14860,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"su",
@@ -13824,6 +15000,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ta",
@@ -13985,6 +15168,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ti",
@@ -14020,6 +15210,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"tr",
@@ -14062,6 +15259,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"uk",
@@ -14139,6 +15343,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"ur",
@@ -14174,6 +15385,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"uz",
@@ -14209,6 +15427,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"vi",
@@ -14433,6 +15658,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"anthropic\/claude-sonnet-4",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"anthropic\/claude-sonnet-4",
     "bcp_47":"yue",
@@ -14769,6 +16001,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ar",
@@ -14881,6 +16120,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"as",
@@ -14916,6 +16162,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"awa",
@@ -14951,6 +16204,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"az",
@@ -14986,6 +16246,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"be",
@@ -15021,6 +16288,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"bho",
@@ -15175,6 +16449,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ceb",
@@ -15210,6 +16491,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ckb",
@@ -15490,6 +16778,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"fa",
@@ -15658,6 +16953,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"gu",
@@ -15742,6 +17044,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"hi",
@@ -15896,6 +17205,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"id",
@@ -15987,6 +17303,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ilo",
@@ -16120,6 +17443,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"jv",
@@ -16190,6 +17520,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"kk",
@@ -16225,6 +17562,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"km",
@@ -16260,6 +17604,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"kn",
@@ -16295,6 +17646,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ko",
@@ -16407,6 +17765,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"mai",
@@ -16442,6 +17807,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"mg",
@@ -16484,6 +17856,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ml",
@@ -16519,6 +17898,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"mr",
@@ -16554,6 +17940,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ms",
@@ -16596,6 +17989,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"my",
@@ -16631,6 +18031,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ne",
@@ -16722,6 +18129,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ny",
@@ -16813,6 +18227,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"or",
@@ -16848,6 +18269,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"pa",
@@ -16925,6 +18353,13 @@
     "metric":"chrf",
     "score":0.5692291394
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"pt",
@@ -16932,6 +18367,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"pt",
@@ -16967,6 +18409,13 @@
     "metric":"chrf",
     "score":0.6547073078
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ro",
@@ -17121,6 +18570,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"sd",
@@ -17156,6 +18612,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"si",
@@ -17247,6 +18710,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"so",
@@ -17289,6 +18759,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"sr",
@@ -17331,6 +18808,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"su",
@@ -17464,6 +18948,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ta",
@@ -17625,6 +19116,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ti",
@@ -17660,6 +19158,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"tr",
@@ -17702,6 +19207,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"uk",
@@ -17779,6 +19291,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"ur",
@@ -17814,6 +19333,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"uz",
@@ -17849,6 +19375,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"vi",
@@ -18073,6 +19606,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"deepseek\/deepseek-chat",
     "bcp_47":"yue",
@@ -18409,6 +19949,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ar",
@@ -18521,6 +20068,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"as",
@@ -18556,6 +20110,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"awa",
@@ -18591,6 +20152,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"az",
@@ -18626,6 +20194,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"be",
@@ -18661,6 +20236,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"bho",
@@ -18815,6 +20397,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ceb",
@@ -18850,6 +20439,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ckb",
@@ -19130,6 +20726,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"fa",
@@ -19298,6 +20901,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"gu",
@@ -19382,6 +20992,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"hi",
@@ -19536,6 +21153,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"id",
@@ -19627,6 +21251,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ilo",
@@ -19760,6 +21391,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"jv",
@@ -19830,6 +21468,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"kk",
@@ -19865,6 +21510,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"km",
@@ -19900,6 +21552,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"kn",
@@ -19935,6 +21594,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ko",
@@ -20047,6 +21713,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"mai",
@@ -20082,6 +21755,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"mg",
@@ -20124,6 +21804,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ml",
@@ -20159,6 +21846,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"mr",
@@ -20194,6 +21888,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ms",
@@ -20236,6 +21937,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"my",
@@ -20271,6 +21979,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ne",
@@ -20362,6 +22077,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ny",
@@ -20453,6 +22175,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"or",
@@ -20488,6 +22217,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pa",
@@ -20565,6 +22301,13 @@
     "metric":"chrf",
     "score":0.540420297
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pt",
@@ -20572,6 +22315,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"pt",
@@ -20607,6 +22357,13 @@
     "metric":"chrf",
     "score":0.6710753294
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ro",
@@ -20761,6 +22518,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"sd",
@@ -20796,6 +22560,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"si",
@@ -20887,6 +22658,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"so",
@@ -20929,6 +22707,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"sr",
@@ -20971,6 +22756,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"su",
@@ -21104,6 +22896,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ta",
@@ -21265,6 +23064,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ti",
@@ -21300,6 +23106,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"tr",
@@ -21342,6 +23155,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"uk",
@@ -21419,6 +23239,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"ur",
@@ -21454,6 +23281,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"uz",
@@ -21489,6 +23323,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"vi",
@@ -21713,6 +23554,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"deepseek\/deepseek-chat-v3-0324",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"deepseek\/deepseek-chat-v3-0324",
     "bcp_47":"yue",
@@ -22049,6 +23897,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ar",
@@ -22161,6 +24016,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"as",
@@ -22196,6 +24058,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"awa",
@@ -22231,6 +24100,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"az",
@@ -22266,6 +24142,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"be",
@@ -22301,6 +24184,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"bho",
@@ -22455,6 +24345,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ceb",
@@ -22490,6 +24387,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ckb",
@@ -22770,6 +24674,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"fa",
@@ -22938,6 +24849,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"gu",
@@ -23022,6 +24940,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"hi",
@@ -23176,6 +25101,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"id",
@@ -23267,6 +25199,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ilo",
@@ -23400,6 +25339,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"jv",
@@ -23470,6 +25416,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"kk",
@@ -23505,6 +25458,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"km",
@@ -23540,6 +25500,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"kn",
@@ -23575,6 +25542,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ko",
@@ -23687,6 +25661,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"mai",
@@ -23722,6 +25703,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"mg",
@@ -23764,6 +25752,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ml",
@@ -23799,6 +25794,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"mr",
@@ -23834,6 +25836,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ms",
@@ -23876,6 +25885,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"my",
@@ -23911,6 +25927,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ne",
@@ -24002,6 +26025,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ny",
@@ -24093,6 +26123,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"or",
@@ -24128,6 +26165,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"pa",
@@ -24205,6 +26249,13 @@
     "metric":"chrf",
     "score":0.4560149918
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"pt",
@@ -24212,6 +26263,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"pt",
@@ -24247,6 +26305,13 @@
     "metric":"chrf",
     "score":0.649967582
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ro",
@@ -24401,6 +26466,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"sd",
@@ -24436,6 +26508,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"si",
@@ -24527,6 +26606,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"so",
@@ -24569,6 +26655,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"sr",
@@ -24611,6 +26704,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"su",
@@ -24744,6 +26844,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ta",
@@ -24905,6 +27012,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ti",
@@ -24940,6 +27054,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"tr",
@@ -24982,6 +27103,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"uk",
@@ -25059,6 +27187,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"ur",
@@ -25094,6 +27229,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"uz",
@@ -25129,6 +27271,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"vi",
@@ -25353,6 +27502,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"deepseek\/deepseek-r1",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"deepseek\/deepseek-r1",
     "bcp_47":"yue",
@@ -27201,6 +29357,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ar",
@@ -27313,6 +29476,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"as",
@@ -27348,6 +29518,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"awa",
@@ -27383,6 +29560,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"az",
@@ -27418,6 +29602,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"be",
@@ -27453,6 +29644,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"bho",
@@ -27607,6 +29805,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ceb",
@@ -27642,6 +29847,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ckb",
@@ -27922,6 +30134,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"fa",
@@ -28090,6 +30309,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"gu",
@@ -28174,6 +30400,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"hi",
@@ -28328,6 +30561,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"id",
@@ -28419,6 +30659,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ilo",
@@ -28552,6 +30799,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"jv",
@@ -28622,6 +30876,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"kk",
@@ -28657,6 +30918,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"km",
@@ -28692,6 +30960,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"kn",
@@ -28727,6 +31002,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ko",
@@ -28839,6 +31121,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"mai",
@@ -28874,6 +31163,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"mg",
@@ -28916,6 +31212,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ml",
@@ -28951,6 +31254,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"mr",
@@ -28986,6 +31296,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ms",
@@ -29028,6 +31345,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"my",
@@ -29063,6 +31387,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ne",
@@ -29154,6 +31485,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ny",
@@ -29245,6 +31583,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"or",
@@ -29280,6 +31625,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pa",
@@ -29357,6 +31709,13 @@
     "metric":"chrf",
     "score":0.579529149
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pt",
@@ -29364,6 +31723,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"pt",
@@ -29399,6 +31765,13 @@
     "metric":"chrf",
     "score":0.6966172136
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ro",
@@ -29553,6 +31926,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"sd",
@@ -29588,6 +31968,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"si",
@@ -29679,6 +32066,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"so",
@@ -29721,6 +32115,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"sr",
@@ -29763,6 +32164,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"su",
@@ -29896,6 +32304,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ta",
@@ -30057,6 +32472,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ti",
@@ -30092,6 +32514,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"tr",
@@ -30134,6 +32563,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"uk",
@@ -30211,6 +32647,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"ur",
@@ -30246,6 +32689,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"uz",
@@ -30281,6 +32731,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"vi",
@@ -30505,6 +32962,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-001",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.0-flash-001",
     "bcp_47":"yue",
@@ -30841,6 +33305,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ar",
@@ -30953,6 +33424,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"as",
@@ -30988,6 +33466,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"awa",
@@ -31023,6 +33508,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"az",
@@ -31058,6 +33550,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"be",
@@ -31093,6 +33592,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"bho",
@@ -31247,6 +33753,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ceb",
@@ -31282,6 +33795,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ckb",
@@ -31562,6 +34082,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"fa",
@@ -31730,6 +34257,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"gu",
@@ -31814,6 +34348,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"hi",
@@ -31968,6 +34509,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"id",
@@ -32059,6 +34607,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ilo",
@@ -32192,6 +34747,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"jv",
@@ -32262,6 +34824,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"kk",
@@ -32297,6 +34866,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"km",
@@ -32332,6 +34908,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"kn",
@@ -32367,6 +34950,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ko",
@@ -32479,6 +35069,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"mai",
@@ -32514,6 +35111,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"mg",
@@ -32556,6 +35160,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ml",
@@ -32591,6 +35202,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"mr",
@@ -32626,6 +35244,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ms",
@@ -32668,6 +35293,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"my",
@@ -32703,6 +35335,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ne",
@@ -32794,6 +35433,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ny",
@@ -32885,6 +35531,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"or",
@@ -32920,6 +35573,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pa",
@@ -32997,6 +35657,13 @@
     "metric":"chrf",
     "score":0.5847615168
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pt",
@@ -33004,6 +35671,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"pt",
@@ -33039,6 +35713,13 @@
     "metric":"chrf",
     "score":0.6588223005
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ro",
@@ -33193,6 +35874,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"sd",
@@ -33228,6 +35916,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"si",
@@ -33319,6 +36014,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"so",
@@ -33361,6 +36063,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"sr",
@@ -33403,6 +36112,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"su",
@@ -33536,6 +36252,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ta",
@@ -33697,6 +36420,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ti",
@@ -33732,6 +36462,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"tr",
@@ -33774,6 +36511,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"uk",
@@ -33851,6 +36595,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"ur",
@@ -33886,6 +36637,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"uz",
@@ -33921,6 +36679,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"vi",
@@ -34145,6 +36910,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.0-flash-lite-001",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.0-flash-lite-001",
     "bcp_47":"yue",
@@ -34446,6 +37218,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ar",
@@ -34558,6 +37337,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"as",
@@ -34593,6 +37379,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"awa",
@@ -34628,6 +37421,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"az",
@@ -34663,6 +37463,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"be",
@@ -34698,6 +37505,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"bho",
@@ -34782,6 +37596,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ceb",
@@ -34817,6 +37638,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ckb",
@@ -35097,6 +37925,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"fa",
@@ -35265,6 +38100,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"gu",
@@ -35349,6 +38191,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"hi",
@@ -35468,6 +38317,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"id",
@@ -35559,6 +38415,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ilo",
@@ -35692,6 +38555,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"jv",
@@ -35727,6 +38597,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"kk",
@@ -35762,6 +38639,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"km",
@@ -35797,6 +38681,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"kn",
@@ -35832,6 +38723,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ko",
@@ -35944,6 +38842,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"mai",
@@ -35979,6 +38884,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"mg",
@@ -36021,6 +38933,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ml",
@@ -36056,6 +38975,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"mr",
@@ -36091,6 +39017,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ms",
@@ -36133,6 +39066,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"my",
@@ -36168,6 +39108,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ne",
@@ -36259,6 +39206,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ny",
@@ -36350,6 +39304,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"or",
@@ -36385,6 +39346,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"pa",
@@ -36462,6 +39430,13 @@
     "metric":"chrf",
     "score":0.5993345379
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"pt",
@@ -36469,6 +39444,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"pt",
@@ -36504,6 +39486,13 @@
     "metric":"chrf",
     "score":0.6755172019
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ro",
@@ -36658,6 +39647,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"sd",
@@ -36693,6 +39689,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"si",
@@ -36784,6 +39787,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"so",
@@ -36826,6 +39836,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"sr",
@@ -36868,6 +39885,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"su",
@@ -37001,6 +40025,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ta",
@@ -37127,6 +40158,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ti",
@@ -37162,6 +40200,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"tr",
@@ -37204,6 +40249,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"uk",
@@ -37246,6 +40298,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"ur",
@@ -37281,6 +40340,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"uz",
@@ -37316,6 +40382,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"vi",
@@ -37540,6 +40613,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-flash",
     "bcp_47":"yue",
@@ -37841,6 +40921,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ar",
@@ -37953,6 +41040,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"as",
@@ -37988,6 +41082,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"awa",
@@ -38023,6 +41124,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"az",
@@ -38058,6 +41166,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"be",
@@ -38093,6 +41208,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"bho",
@@ -38177,6 +41299,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ceb",
@@ -38212,6 +41341,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ckb",
@@ -38492,6 +41628,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"fa",
@@ -38660,6 +41803,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"gu",
@@ -38744,6 +41894,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"hi",
@@ -38863,6 +42020,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"id",
@@ -38954,6 +42118,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ilo",
@@ -39087,6 +42258,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"jv",
@@ -39122,6 +42300,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"kk",
@@ -39157,6 +42342,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"km",
@@ -39192,6 +42384,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"kn",
@@ -39227,6 +42426,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ko",
@@ -39339,6 +42545,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"mai",
@@ -39374,6 +42587,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"mg",
@@ -39416,6 +42636,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ml",
@@ -39451,6 +42678,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"mr",
@@ -39486,6 +42720,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ms",
@@ -39528,6 +42769,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"my",
@@ -39563,6 +42811,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ne",
@@ -39654,6 +42909,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ny",
@@ -39745,6 +43007,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"or",
@@ -39780,6 +43049,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"pa",
@@ -39857,6 +43133,13 @@
     "metric":"chrf",
     "score":0.5954877296
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"pt",
@@ -39864,6 +43147,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"pt",
@@ -39899,6 +43189,13 @@
     "metric":"chrf",
     "score":0.7255352889
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ro",
@@ -40053,6 +43350,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"sd",
@@ -40088,6 +43392,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"si",
@@ -40179,6 +43490,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"so",
@@ -40221,6 +43539,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"sr",
@@ -40263,6 +43588,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"su",
@@ -40396,6 +43728,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ta",
@@ -40522,6 +43861,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ti",
@@ -40557,6 +43903,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"tr",
@@ -40599,6 +43952,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"uk",
@@ -40641,6 +44001,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"ur",
@@ -40676,6 +44043,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"uz",
@@ -40711,6 +44085,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"vi",
@@ -40935,6 +44316,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-lite-preview-06-17",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-flash-lite-preview-06-17",
     "bcp_47":"yue",
@@ -41271,6 +44659,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ar",
@@ -41383,6 +44778,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"as",
@@ -41418,6 +44820,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"awa",
@@ -41453,6 +44862,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"az",
@@ -41488,6 +44904,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"be",
@@ -41523,6 +44946,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"bho",
@@ -41677,6 +45107,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ceb",
@@ -41712,6 +45149,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ckb",
@@ -41992,6 +45436,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"fa",
@@ -42160,6 +45611,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"gu",
@@ -42244,6 +45702,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"hi",
@@ -42398,6 +45863,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"id",
@@ -42489,6 +45961,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ilo",
@@ -42622,6 +46101,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"jv",
@@ -42692,6 +46178,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"kk",
@@ -42727,6 +46220,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"km",
@@ -42762,6 +46262,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"kn",
@@ -42797,6 +46304,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ko",
@@ -42909,6 +46423,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"mai",
@@ -42944,6 +46465,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"mg",
@@ -42986,6 +46514,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ml",
@@ -43021,6 +46556,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"mr",
@@ -43056,6 +46598,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ms",
@@ -43098,6 +46647,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"my",
@@ -43133,6 +46689,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ne",
@@ -43224,6 +46787,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ny",
@@ -43315,6 +46885,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"or",
@@ -43350,6 +46927,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pa",
@@ -43427,6 +47011,13 @@
     "metric":"chrf",
     "score":0.6077722316
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pt",
@@ -43434,6 +47025,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"pt",
@@ -43469,6 +47067,13 @@
     "metric":"chrf",
     "score":0.7045563325
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ro",
@@ -43623,6 +47228,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"sd",
@@ -43658,6 +47270,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"si",
@@ -43749,6 +47368,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"so",
@@ -43791,6 +47417,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"sr",
@@ -43833,6 +47466,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"su",
@@ -43966,6 +47606,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ta",
@@ -44127,6 +47774,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ti",
@@ -44162,6 +47816,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"tr",
@@ -44204,6 +47865,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"uk",
@@ -44281,6 +47949,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"ur",
@@ -44316,6 +47991,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"uz",
@@ -44351,6 +48033,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"vi",
@@ -44575,6 +48264,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-flash-preview",
     "bcp_47":"yue",
@@ -44911,6 +48607,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ar",
@@ -45023,6 +48726,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"as",
@@ -45058,6 +48768,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"awa",
@@ -45093,6 +48810,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"az",
@@ -45128,6 +48852,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"be",
@@ -45163,6 +48894,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"bho",
@@ -45317,6 +49055,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ceb",
@@ -45352,6 +49097,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ckb",
@@ -45632,6 +49384,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"fa",
@@ -45800,6 +49559,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"gu",
@@ -45884,6 +49650,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"hi",
@@ -46038,6 +49811,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"id",
@@ -46129,6 +49909,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ilo",
@@ -46262,6 +50049,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"jv",
@@ -46332,6 +50126,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"kk",
@@ -46367,6 +50168,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"km",
@@ -46402,6 +50210,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"kn",
@@ -46437,6 +50252,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ko",
@@ -46549,6 +50371,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"mai",
@@ -46584,6 +50413,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"mg",
@@ -46626,6 +50462,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ml",
@@ -46661,6 +50504,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"mr",
@@ -46696,6 +50546,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ms",
@@ -46738,6 +50595,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"my",
@@ -46773,6 +50637,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ne",
@@ -46864,6 +50735,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ny",
@@ -46955,6 +50833,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"or",
@@ -46990,6 +50875,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pa",
@@ -47067,6 +50959,13 @@
     "metric":"chrf",
     "score":0.5940500139
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pt",
@@ -47074,6 +50973,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"pt",
@@ -47109,6 +51015,13 @@
     "metric":"chrf",
     "score":0.6891420805
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ro",
@@ -47263,6 +51176,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"sd",
@@ -47298,6 +51218,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"si",
@@ -47389,6 +51316,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"so",
@@ -47431,6 +51365,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"sr",
@@ -47473,6 +51414,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"su",
@@ -47606,6 +51554,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ta",
@@ -47767,6 +51722,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ti",
@@ -47802,6 +51764,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"tr",
@@ -47844,6 +51813,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"uk",
@@ -47921,6 +51897,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"ur",
@@ -47956,6 +51939,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"uz",
@@ -47991,6 +51981,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"vi",
@@ -48215,6 +52212,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-2.5-flash-preview-05-20",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-flash-preview-05-20",
     "bcp_47":"yue",
@@ -50791,6 +54795,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ar",
@@ -50903,6 +54914,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"as",
@@ -50938,6 +54956,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"awa",
@@ -50973,6 +54998,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"az",
@@ -51008,6 +55040,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"be",
@@ -51043,6 +55082,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"bho",
@@ -51197,6 +55243,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ceb",
@@ -51232,6 +55285,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ckb",
@@ -51512,6 +55572,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"fa",
@@ -51680,6 +55747,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"gu",
@@ -51764,6 +55838,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"hi",
@@ -51918,6 +55999,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"id",
@@ -52009,6 +56097,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ilo",
@@ -52142,6 +56237,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"jv",
@@ -52212,6 +56314,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"kk",
@@ -52247,6 +56356,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"km",
@@ -52282,6 +56398,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"kn",
@@ -52317,6 +56440,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ko",
@@ -52429,6 +56559,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"mai",
@@ -52464,6 +56601,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"mg",
@@ -52506,6 +56650,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ml",
@@ -52541,6 +56692,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"mr",
@@ -52576,6 +56734,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ms",
@@ -52618,6 +56783,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"my",
@@ -52653,6 +56825,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ne",
@@ -52744,6 +56923,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ny",
@@ -52835,6 +57021,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"or",
@@ -52870,6 +57063,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"pa",
@@ -52947,6 +57147,13 @@
     "metric":"chrf",
     "score":0.1925827805
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"pt",
@@ -52954,6 +57161,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"pt",
@@ -52989,6 +57203,13 @@
     "metric":"chrf",
     "score":0.2615431845
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ro",
@@ -53143,6 +57364,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"sd",
@@ -53178,6 +57406,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"si",
@@ -53269,6 +57504,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"so",
@@ -53311,6 +57553,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"sr",
@@ -53353,6 +57602,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"su",
@@ -53486,6 +57742,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ta",
@@ -53647,6 +57910,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ti",
@@ -53682,6 +57952,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"tr",
@@ -53724,6 +58001,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"uk",
@@ -53801,6 +58085,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"ur",
@@ -53836,6 +58127,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"uz",
@@ -53871,6 +58169,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"vi",
@@ -54095,6 +58400,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"google\/gemini-2.5-pro-preview-05-06",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-2.5-pro-preview-05-06",
     "bcp_47":"yue",
@@ -54431,6 +58743,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ar",
@@ -54543,6 +58862,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"as",
@@ -54578,6 +58904,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"awa",
@@ -54613,6 +58946,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"az",
@@ -54648,6 +58988,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"be",
@@ -54683,6 +59030,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"bho",
@@ -54837,6 +59191,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ceb",
@@ -54872,6 +59233,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ckb",
@@ -55152,6 +59520,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"fa",
@@ -55320,6 +59695,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"gu",
@@ -55404,6 +59786,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"hi",
@@ -55558,6 +59947,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"id",
@@ -55649,6 +60045,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ilo",
@@ -55782,6 +60185,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"jv",
@@ -55852,6 +60262,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"kk",
@@ -55887,6 +60304,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"km",
@@ -55922,6 +60346,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"kn",
@@ -55957,6 +60388,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ko",
@@ -56069,6 +60507,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"mai",
@@ -56104,6 +60549,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"mg",
@@ -56146,6 +60598,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ml",
@@ -56181,6 +60640,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"mr",
@@ -56216,6 +60682,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ms",
@@ -56258,6 +60731,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"my",
@@ -56293,6 +60773,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ne",
@@ -56384,6 +60871,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ny",
@@ -56475,6 +60969,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"or",
@@ -56510,6 +61011,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"pa",
@@ -56587,6 +61095,13 @@
     "metric":"chrf",
     "score":0.6320824157
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"pt",
@@ -56594,6 +61109,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"pt",
@@ -56629,6 +61151,13 @@
     "metric":"chrf",
     "score":0.6844540285
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ro",
@@ -56783,6 +61312,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"sd",
@@ -56818,6 +61354,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"si",
@@ -56909,6 +61452,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"so",
@@ -56951,6 +61501,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"sr",
@@ -56993,6 +61550,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"su",
@@ -57126,6 +61690,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ta",
@@ -57287,6 +61858,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ti",
@@ -57322,6 +61900,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"tr",
@@ -57364,6 +61949,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"uk",
@@ -57441,6 +62033,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"ur",
@@ -57476,6 +62075,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"uz",
@@ -57511,6 +62117,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"vi",
@@ -57735,6 +62348,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-flash-1.5",
     "bcp_47":"yue",
@@ -58071,6 +62691,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ar",
@@ -58183,6 +62810,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"as",
@@ -58218,6 +62852,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"awa",
@@ -58253,6 +62894,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"az",
@@ -58288,6 +62936,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"be",
@@ -58323,6 +62978,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"bho",
@@ -58477,6 +63139,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ceb",
@@ -58512,6 +63181,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ckb",
@@ -58792,6 +63468,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"fa",
@@ -58960,6 +63643,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"gu",
@@ -59044,6 +63734,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"hi",
@@ -59198,6 +63895,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"id",
@@ -59289,6 +63993,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ilo",
@@ -59422,6 +64133,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"jv",
@@ -59492,6 +64210,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"kk",
@@ -59527,6 +64252,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"km",
@@ -59562,6 +64294,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"kn",
@@ -59597,6 +64336,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ko",
@@ -59709,6 +64455,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"mai",
@@ -59744,6 +64497,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"mg",
@@ -59786,6 +64546,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ml",
@@ -59821,6 +64588,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"mr",
@@ -59856,6 +64630,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ms",
@@ -59898,6 +64679,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"my",
@@ -59933,6 +64721,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ne",
@@ -60024,6 +64819,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ny",
@@ -60115,6 +64917,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"or",
@@ -60150,6 +64959,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pa",
@@ -60227,6 +65043,13 @@
     "metric":"chrf",
     "score":0.5924241261
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pt",
@@ -60234,6 +65057,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"pt",
@@ -60269,6 +65099,13 @@
     "metric":"chrf",
     "score":0.6509885745
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ro",
@@ -60423,6 +65260,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"sd",
@@ -60458,6 +65302,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"si",
@@ -60549,6 +65400,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"so",
@@ -60591,6 +65449,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"sr",
@@ -60633,6 +65498,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"su",
@@ -60766,6 +65638,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ta",
@@ -60927,6 +65806,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ti",
@@ -60962,6 +65848,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"tr",
@@ -61004,6 +65897,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"uk",
@@ -61081,6 +65981,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"ur",
@@ -61116,6 +66023,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"uz",
@@ -61151,6 +66065,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"vi",
@@ -61375,6 +66296,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemini-flash-1.5-8b",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"google\/gemini-flash-1.5-8b",
     "bcp_47":"yue",
@@ -61711,6 +66639,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ar",
@@ -61823,6 +66758,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"as",
@@ -61858,6 +66800,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"awa",
@@ -61893,6 +66842,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"az",
@@ -61928,6 +66884,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"be",
@@ -61963,6 +66926,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"bho",
@@ -62117,6 +67087,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ceb",
@@ -62152,6 +67129,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ckb",
@@ -62432,6 +67416,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"fa",
@@ -62600,6 +67591,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"gu",
@@ -62684,6 +67682,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"hi",
@@ -62838,6 +67843,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"id",
@@ -62929,6 +67941,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ilo",
@@ -63062,6 +68081,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"jv",
@@ -63132,6 +68158,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"kk",
@@ -63167,6 +68200,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"km",
@@ -63202,6 +68242,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"kn",
@@ -63237,6 +68284,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ko",
@@ -63349,6 +68403,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"mai",
@@ -63384,6 +68445,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"mg",
@@ -63426,6 +68494,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ml",
@@ -63461,6 +68536,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"mr",
@@ -63496,6 +68578,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ms",
@@ -63538,6 +68627,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"my",
@@ -63573,6 +68669,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ne",
@@ -63664,6 +68767,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ny",
@@ -63755,6 +68865,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"or",
@@ -63790,6 +68907,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"pa",
@@ -63867,6 +68991,13 @@
     "metric":"chrf",
     "score":0.5891983505
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"pt",
@@ -63874,6 +69005,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"pt",
@@ -63909,6 +69047,13 @@
     "metric":"chrf",
     "score":0.6903236014
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ro",
@@ -64063,6 +69208,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"sd",
@@ -64098,6 +69250,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"si",
@@ -64189,6 +69348,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"so",
@@ -64231,6 +69397,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"sr",
@@ -64273,6 +69446,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"su",
@@ -64406,6 +69586,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ta",
@@ -64567,6 +69754,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ti",
@@ -64602,6 +69796,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"tr",
@@ -64644,6 +69845,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"uk",
@@ -64721,6 +69929,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"ur",
@@ -64756,6 +69971,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"uz",
@@ -64791,6 +70013,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"vi",
@@ -65015,6 +70244,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"google\/gemma-3-27b-it",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"google\/gemma-3-27b-it",
     "bcp_47":"yue",
@@ -67647,6 +72883,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ar",
@@ -67759,6 +73002,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"as",
@@ -67794,6 +73044,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"awa",
@@ -67829,6 +73086,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"az",
@@ -67864,6 +73128,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"be",
@@ -67899,6 +73170,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"bho",
@@ -68053,6 +73331,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ceb",
@@ -68088,6 +73373,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ckb",
@@ -68368,6 +73660,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"fa",
@@ -68536,6 +73835,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"gu",
@@ -68620,6 +73926,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"hi",
@@ -68774,6 +74087,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"id",
@@ -68865,6 +74185,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ilo",
@@ -68998,6 +74325,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"jv",
@@ -69068,6 +74402,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"kk",
@@ -69103,6 +74444,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"km",
@@ -69138,6 +74486,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"kn",
@@ -69173,6 +74528,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ko",
@@ -69285,6 +74647,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"mai",
@@ -69320,6 +74689,13 @@
     "metric":"accuracy",
     "score":0.3
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"mg",
@@ -69362,6 +74738,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ml",
@@ -69397,6 +74780,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"mr",
@@ -69432,6 +74822,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ms",
@@ -69474,6 +74871,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"my",
@@ -69509,6 +74913,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ne",
@@ -69600,6 +75011,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ny",
@@ -69691,6 +75109,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"or",
@@ -69726,6 +75151,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pa",
@@ -69803,6 +75235,13 @@
     "metric":"chrf",
     "score":0.3907504991
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pt",
@@ -69810,6 +75249,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"pt",
@@ -69845,6 +75291,13 @@
     "metric":"chrf",
     "score":0.4401829864
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ro",
@@ -69999,6 +75452,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"sd",
@@ -70034,6 +75494,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"si",
@@ -70125,6 +75592,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"so",
@@ -70167,6 +75641,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"sr",
@@ -70209,6 +75690,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"su",
@@ -70342,6 +75830,13 @@
     "metric":"accuracy",
     "score":0.3
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ta",
@@ -70503,6 +75998,13 @@
     "metric":"accuracy",
     "score":0.3
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ti",
@@ -70538,6 +76040,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"tr",
@@ -70580,6 +76089,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"uk",
@@ -70657,6 +76173,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"ur",
@@ -70692,6 +76215,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"uz",
@@ -70727,6 +76257,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"vi",
@@ -70951,6 +76488,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"gryphe\/mythomax-l2-13b",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"gryphe\/mythomax-l2-13b",
     "bcp_47":"yue",
@@ -71287,6 +76831,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ar",
@@ -71399,6 +76950,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"as",
@@ -71434,6 +76992,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"awa",
@@ -71469,6 +77034,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"az",
@@ -71504,6 +77076,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"be",
@@ -71539,6 +77118,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"bho",
@@ -71693,6 +77279,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ceb",
@@ -71728,6 +77321,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ckb",
@@ -72008,6 +77608,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"fa",
@@ -72176,6 +77783,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"gu",
@@ -72260,6 +77874,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"hi",
@@ -72414,6 +78035,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"id",
@@ -72505,6 +78133,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ilo",
@@ -72638,6 +78273,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"jv",
@@ -72708,6 +78350,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"kk",
@@ -72743,6 +78392,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"km",
@@ -72778,6 +78434,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"kn",
@@ -72813,6 +78476,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ko",
@@ -72925,6 +78595,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"mai",
@@ -72960,6 +78637,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"mg",
@@ -73002,6 +78686,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ml",
@@ -73037,6 +78728,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"mr",
@@ -73072,6 +78770,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ms",
@@ -73114,6 +78819,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"my",
@@ -73149,6 +78861,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ne",
@@ -73240,6 +78959,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ny",
@@ -73331,6 +79057,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"or",
@@ -73366,6 +79099,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pa",
@@ -73443,6 +79183,13 @@
     "metric":"chrf",
     "score":0.5541703282
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pt",
@@ -73450,6 +79197,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"pt",
@@ -73485,6 +79239,13 @@
     "metric":"chrf",
     "score":0.6132478102
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ro",
@@ -73639,6 +79400,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"sd",
@@ -73674,6 +79442,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"si",
@@ -73765,6 +79540,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"so",
@@ -73807,6 +79589,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"sr",
@@ -73849,6 +79638,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"su",
@@ -73982,6 +79778,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ta",
@@ -74143,6 +79946,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ti",
@@ -74178,6 +79988,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"tr",
@@ -74220,6 +80037,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"uk",
@@ -74297,6 +80121,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"ur",
@@ -74332,6 +80163,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"uz",
@@ -74367,6 +80205,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"vi",
@@ -74591,6 +80436,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3-70b-instruct",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3-70b-instruct",
     "bcp_47":"yue",
@@ -74927,6 +80779,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ar",
@@ -75039,6 +80898,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"as",
@@ -75074,6 +80940,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"awa",
@@ -75109,6 +80982,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"az",
@@ -75144,6 +81024,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"be",
@@ -75179,6 +81066,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"bho",
@@ -75333,6 +81227,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ceb",
@@ -75368,6 +81269,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ckb",
@@ -75648,6 +81556,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"fa",
@@ -75816,6 +81731,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"gu",
@@ -75900,6 +81822,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"hi",
@@ -76054,6 +81983,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"id",
@@ -76145,6 +82081,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ilo",
@@ -76278,6 +82221,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"jv",
@@ -76348,6 +82298,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"kk",
@@ -76383,6 +82340,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"km",
@@ -76418,6 +82382,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"kn",
@@ -76453,6 +82424,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ko",
@@ -76565,6 +82543,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"mai",
@@ -76600,6 +82585,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"mg",
@@ -76642,6 +82634,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ml",
@@ -76677,6 +82676,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"mr",
@@ -76712,6 +82718,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ms",
@@ -76754,6 +82767,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"my",
@@ -76789,6 +82809,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ne",
@@ -76880,6 +82907,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ny",
@@ -76971,6 +83005,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"or",
@@ -77006,6 +83047,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"pa",
@@ -77083,6 +83131,13 @@
     "metric":"chrf",
     "score":0.5569354008
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"pt",
@@ -77090,6 +83145,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"pt",
@@ -77125,6 +83187,13 @@
     "metric":"chrf",
     "score":0.6761551234
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ro",
@@ -77279,6 +83348,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"sd",
@@ -77314,6 +83390,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"si",
@@ -77405,6 +83488,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"so",
@@ -77447,6 +83537,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"sr",
@@ -77489,6 +83586,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"su",
@@ -77622,6 +83726,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ta",
@@ -77783,6 +83894,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ti",
@@ -77818,6 +83936,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"tr",
@@ -77860,6 +83985,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"uk",
@@ -77937,6 +84069,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"ur",
@@ -77972,6 +84111,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"uz",
@@ -78007,6 +84153,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"vi",
@@ -78231,6 +84384,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.1-70b-instruct",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.1-70b-instruct",
     "bcp_47":"yue",
@@ -83047,6 +89207,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ar",
@@ -83159,6 +89326,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"as",
@@ -83194,6 +89368,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"awa",
@@ -83229,6 +89410,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"az",
@@ -83264,6 +89452,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"be",
@@ -83299,6 +89494,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"bho",
@@ -83453,6 +89655,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ceb",
@@ -83488,6 +89697,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ckb",
@@ -83768,6 +89984,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"fa",
@@ -83936,6 +90159,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"gu",
@@ -84020,6 +90250,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"hi",
@@ -84174,6 +90411,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"id",
@@ -84265,6 +90509,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ilo",
@@ -84398,6 +90649,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"jv",
@@ -84468,6 +90726,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"kk",
@@ -84503,6 +90768,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"km",
@@ -84538,6 +90810,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"kn",
@@ -84573,6 +90852,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ko",
@@ -84685,6 +90971,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"mai",
@@ -84720,6 +91013,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"mg",
@@ -84762,6 +91062,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ml",
@@ -84797,6 +91104,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"mr",
@@ -84832,6 +91146,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ms",
@@ -84874,6 +91195,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"my",
@@ -84909,6 +91237,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ne",
@@ -85000,6 +91335,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ny",
@@ -85091,6 +91433,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"or",
@@ -85126,6 +91475,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pa",
@@ -85203,6 +91559,13 @@
     "metric":"chrf",
     "score":0.5550868321
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pt",
@@ -85210,6 +91573,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"pt",
@@ -85245,6 +91615,13 @@
     "metric":"chrf",
     "score":0.6712742861
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ro",
@@ -85399,6 +91776,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"sd",
@@ -85434,6 +91818,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"si",
@@ -85525,6 +91916,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"so",
@@ -85567,6 +91965,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"sr",
@@ -85609,6 +92014,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"su",
@@ -85742,6 +92154,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ta",
@@ -85903,6 +92322,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ti",
@@ -85938,6 +92364,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"tr",
@@ -85980,6 +92413,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"uk",
@@ -86057,6 +92497,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"ur",
@@ -86092,6 +92539,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"uz",
@@ -86127,6 +92581,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"vi",
@@ -86351,6 +92812,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-3.3-70b-instruct",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-3.3-70b-instruct",
     "bcp_47":"yue",
@@ -86687,6 +93155,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ar",
@@ -86799,6 +93274,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"as",
@@ -86834,6 +93316,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"awa",
@@ -86869,6 +93358,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"az",
@@ -86904,6 +93400,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"be",
@@ -86939,6 +93442,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"bho",
@@ -87093,6 +93603,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ceb",
@@ -87128,6 +93645,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ckb",
@@ -87408,6 +93932,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"fa",
@@ -87576,6 +94107,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"gu",
@@ -87660,6 +94198,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"hi",
@@ -87814,6 +94359,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"id",
@@ -87905,6 +94457,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ilo",
@@ -88038,6 +94597,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"jv",
@@ -88108,6 +94674,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"kk",
@@ -88143,6 +94716,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"km",
@@ -88178,6 +94758,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"kn",
@@ -88213,6 +94800,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ko",
@@ -88325,6 +94919,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"mai",
@@ -88360,6 +94961,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"mg",
@@ -88402,6 +95010,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ml",
@@ -88437,6 +95052,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"mr",
@@ -88472,6 +95094,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ms",
@@ -88514,6 +95143,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"my",
@@ -88549,6 +95185,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ne",
@@ -88640,6 +95283,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ny",
@@ -88731,6 +95381,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"or",
@@ -88766,6 +95423,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pa",
@@ -88843,6 +95507,13 @@
     "metric":"chrf",
     "score":0.5407300006
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pt",
@@ -88850,6 +95521,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"pt",
@@ -88885,6 +95563,13 @@
     "metric":"chrf",
     "score":0.6491023878
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ro",
@@ -89039,6 +95724,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"sd",
@@ -89074,6 +95766,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"si",
@@ -89165,6 +95864,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"so",
@@ -89207,6 +95913,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"sr",
@@ -89249,6 +95962,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"su",
@@ -89382,6 +96102,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ta",
@@ -89543,6 +96270,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ti",
@@ -89578,6 +96312,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"tr",
@@ -89620,6 +96361,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"uk",
@@ -89697,6 +96445,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"ur",
@@ -89732,6 +96487,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"uz",
@@ -89767,6 +96529,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"vi",
@@ -89991,6 +96760,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"meta-llama\/llama-4-maverick",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"meta-llama\/llama-4-maverick",
     "bcp_47":"yue",
@@ -90327,6 +97103,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ar",
@@ -90439,6 +97222,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"as",
@@ -90474,6 +97264,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"awa",
@@ -90509,6 +97306,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"az",
@@ -90544,6 +97348,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"be",
@@ -90579,6 +97390,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"bho",
@@ -90733,6 +97551,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ceb",
@@ -90768,6 +97593,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ckb",
@@ -91048,6 +97880,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"fa",
@@ -91216,6 +98055,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"gu",
@@ -91300,6 +98146,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"hi",
@@ -91454,6 +98307,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"id",
@@ -91545,6 +98405,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ilo",
@@ -91678,6 +98545,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"jv",
@@ -91748,6 +98622,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"kk",
@@ -91783,6 +98664,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"km",
@@ -91818,6 +98706,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"kn",
@@ -91853,6 +98748,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ko",
@@ -91965,6 +98867,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"mai",
@@ -92000,6 +98909,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"mg",
@@ -92042,6 +98958,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ml",
@@ -92077,6 +99000,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"mr",
@@ -92112,6 +99042,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ms",
@@ -92154,6 +99091,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"my",
@@ -92189,6 +99133,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ne",
@@ -92280,6 +99231,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ny",
@@ -92371,6 +99329,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"or",
@@ -92406,6 +99371,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"pa",
@@ -92483,6 +99455,13 @@
     "metric":"chrf",
     "score":0.4754992095
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"pt",
@@ -92490,6 +99469,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"pt",
@@ -92525,6 +99511,13 @@
     "metric":"chrf",
     "score":0.5827805827
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ro",
@@ -92679,6 +99672,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"sd",
@@ -92714,6 +99714,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"si",
@@ -92805,6 +99812,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"so",
@@ -92847,6 +99861,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"sr",
@@ -92889,6 +99910,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"su",
@@ -93022,6 +100050,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ta",
@@ -93183,6 +100218,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ti",
@@ -93218,6 +100260,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"tr",
@@ -93260,6 +100309,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"uk",
@@ -93337,6 +100393,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"ur",
@@ -93372,6 +100435,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"uz",
@@ -93407,6 +100477,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"vi",
@@ -93631,6 +100708,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4",
     "bcp_47":"yue",
@@ -93967,6 +101051,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ar",
@@ -94079,6 +101170,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"as",
@@ -94114,6 +101212,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"awa",
@@ -94149,6 +101254,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"az",
@@ -94184,6 +101296,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"be",
@@ -94219,6 +101338,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"bho",
@@ -94373,6 +101499,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ceb",
@@ -94408,6 +101541,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ckb",
@@ -94688,6 +101828,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"fa",
@@ -94856,6 +102003,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"gu",
@@ -94940,6 +102094,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"hi",
@@ -95094,6 +102255,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"id",
@@ -95185,6 +102353,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ilo",
@@ -95318,6 +102493,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"jv",
@@ -95388,6 +102570,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"kk",
@@ -95423,6 +102612,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"km",
@@ -95458,6 +102654,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"kn",
@@ -95493,6 +102696,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ko",
@@ -95605,6 +102815,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"mai",
@@ -95640,6 +102857,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"mg",
@@ -95682,6 +102906,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ml",
@@ -95717,6 +102948,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"mr",
@@ -95752,6 +102990,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ms",
@@ -95794,6 +103039,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"my",
@@ -95829,6 +103081,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ne",
@@ -95920,6 +103179,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ny",
@@ -96011,6 +103277,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"or",
@@ -96046,6 +103319,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"pa",
@@ -96123,6 +103403,13 @@
     "metric":"chrf",
     "score":0.2865629267
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"pt",
@@ -96130,6 +103417,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"pt",
@@ -96165,6 +103459,13 @@
     "metric":"chrf",
     "score":0.3618255907
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ro",
@@ -96319,6 +103620,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"sd",
@@ -96354,6 +103662,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"si",
@@ -96445,6 +103760,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"so",
@@ -96487,6 +103809,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"sr",
@@ -96529,6 +103858,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"su",
@@ -96662,6 +103998,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ta",
@@ -96823,6 +104166,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ti",
@@ -96858,6 +104208,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"tr",
@@ -96900,6 +104257,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"uk",
@@ -96977,6 +104341,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"ur",
@@ -97012,6 +104383,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"uz",
@@ -97047,6 +104425,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"vi",
@@ -97271,6 +104656,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"microsoft\/phi-4-multimodal-instruct",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"microsoft\/phi-4-multimodal-instruct",
     "bcp_47":"yue",
@@ -97607,6 +104999,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ar",
@@ -97719,6 +105118,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"as",
@@ -97754,6 +105160,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"awa",
@@ -97789,6 +105202,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"az",
@@ -97824,6 +105244,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"be",
@@ -97859,6 +105286,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"bho",
@@ -98013,6 +105447,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ceb",
@@ -98048,6 +105489,13 @@
     "metric":"accuracy",
     "score":0.5
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ckb",
@@ -98328,6 +105776,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"fa",
@@ -98496,6 +105951,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"gu",
@@ -98580,6 +106042,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"hi",
@@ -98734,6 +106203,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"id",
@@ -98825,6 +106301,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ilo",
@@ -98958,6 +106441,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"jv",
@@ -99028,6 +106518,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"kk",
@@ -99063,6 +106560,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"km",
@@ -99098,6 +106602,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"kn",
@@ -99133,6 +106644,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ko",
@@ -99245,6 +106763,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"mai",
@@ -99280,6 +106805,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"mg",
@@ -99322,6 +106854,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ml",
@@ -99357,6 +106896,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"mr",
@@ -99392,6 +106938,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ms",
@@ -99434,6 +106987,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"my",
@@ -99469,6 +107029,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ne",
@@ -99560,6 +107127,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ny",
@@ -99651,6 +107225,13 @@
     "metric":"accuracy",
     "score":0.0
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"or",
@@ -99686,6 +107267,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"pa",
@@ -99763,6 +107351,13 @@
     "metric":"chrf",
     "score":0.4488014348
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"pt",
@@ -99770,6 +107365,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"pt",
@@ -99805,6 +107407,13 @@
     "metric":"chrf",
     "score":0.5835851988
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ro",
@@ -99959,6 +107568,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"sd",
@@ -99994,6 +107610,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"si",
@@ -100085,6 +107708,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"so",
@@ -100127,6 +107757,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"sr",
@@ -100169,6 +107806,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"su",
@@ -100302,6 +107946,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ta",
@@ -100463,6 +108114,13 @@
     "metric":"accuracy",
     "score":0.1
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ti",
@@ -100498,6 +108156,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"tr",
@@ -100540,6 +108205,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"uk",
@@ -100617,6 +108289,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"ur",
@@ -100652,6 +108331,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"uz",
@@ -100687,6 +108373,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"vi",
@@ -100911,6 +108604,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-nemo",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-nemo",
     "bcp_47":"yue",
@@ -101247,6 +108947,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ar",
@@ -101359,6 +109066,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"as",
@@ -101394,6 +109108,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"awa",
@@ -101429,6 +109150,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"az",
@@ -101464,6 +109192,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"be",
@@ -101499,6 +109234,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"bho",
@@ -101653,6 +109395,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ceb",
@@ -101688,6 +109437,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ckb",
@@ -101968,6 +109724,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"fa",
@@ -102136,6 +109899,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"gu",
@@ -102220,6 +109990,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"hi",
@@ -102374,6 +110151,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"id",
@@ -102465,6 +110249,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ilo",
@@ -102598,6 +110389,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"jv",
@@ -102668,6 +110466,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"kk",
@@ -102703,6 +110508,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"km",
@@ -102738,6 +110550,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"kn",
@@ -102773,6 +110592,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ko",
@@ -102885,6 +110711,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"mai",
@@ -102920,6 +110753,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"mg",
@@ -102962,6 +110802,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ml",
@@ -102997,6 +110844,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"mr",
@@ -103032,6 +110886,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ms",
@@ -103074,6 +110935,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"my",
@@ -103109,6 +110977,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ne",
@@ -103200,6 +111075,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ny",
@@ -103291,6 +111173,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"or",
@@ -103326,6 +111215,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"pa",
@@ -103403,6 +111299,13 @@
     "metric":"chrf",
     "score":0.525989117
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"pt",
@@ -103410,6 +111313,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"pt",
@@ -103445,6 +111355,13 @@
     "metric":"chrf",
     "score":0.6843169799
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ro",
@@ -103599,6 +111516,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"sd",
@@ -103634,6 +111558,13 @@
     "metric":"accuracy",
     "score":0.4
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"si",
@@ -103725,6 +111656,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"so",
@@ -103767,6 +111705,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"sr",
@@ -103809,6 +111754,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"su",
@@ -103942,6 +111894,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ta",
@@ -104103,6 +112062,13 @@
     "metric":"accuracy",
     "score":0.2
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ti",
@@ -104138,6 +112104,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"tr",
@@ -104180,6 +112153,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"uk",
@@ -104257,6 +112237,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"ur",
@@ -104292,6 +112279,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"uz",
@@ -104327,6 +112321,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"vi",
@@ -104551,6 +112552,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-saba",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-saba",
     "bcp_47":"yue",
@@ -104887,6 +112895,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ar",
@@ -104999,6 +113014,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"as",
@@ -105034,6 +113056,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"awa",
@@ -105069,6 +113098,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"az",
@@ -105104,6 +113140,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"be",
@@ -105139,6 +113182,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"bho",
@@ -105293,6 +113343,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ceb",
@@ -105328,6 +113385,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ckb",
@@ -105608,6 +113672,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"fa",
@@ -105776,6 +113847,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"gu",
@@ -105860,6 +113938,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"hi",
@@ -106014,6 +114099,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"id",
@@ -106105,6 +114197,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ilo",
@@ -106238,6 +114337,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"jv",
@@ -106308,6 +114414,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"kk",
@@ -106343,6 +114456,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"km",
@@ -106378,6 +114498,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"kn",
@@ -106413,6 +114540,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ko",
@@ -106525,6 +114659,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"mai",
@@ -106560,6 +114701,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"mg",
@@ -106602,6 +114750,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ml",
@@ -106637,6 +114792,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"mr",
@@ -106672,6 +114834,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ms",
@@ -106714,6 +114883,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"my",
@@ -106749,6 +114925,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ne",
@@ -106840,6 +115023,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ny",
@@ -106931,6 +115121,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"or",
@@ -106966,6 +115163,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"pa",
@@ -107043,6 +115247,13 @@
     "metric":"chrf",
     "score":0.4948927457
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"pt",
@@ -107050,6 +115261,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"pt",
@@ -107085,6 +115303,13 @@
     "metric":"chrf",
     "score":0.6102742767
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ro",
@@ -107239,6 +115464,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"sd",
@@ -107274,6 +115506,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"si",
@@ -107365,6 +115604,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"so",
@@ -107407,6 +115653,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"sr",
@@ -107449,6 +115702,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"su",
@@ -107582,6 +115842,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ta",
@@ -107743,6 +116010,13 @@
     "metric":"accuracy",
     "score":0.3
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ti",
@@ -107778,6 +116052,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"tr",
@@ -107820,6 +116101,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"uk",
@@ -107897,6 +116185,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"ur",
@@ -107932,6 +116227,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"uz",
@@ -107967,6 +116269,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"vi",
@@ -108191,6 +116500,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"mistralai\/mistral-small-3.1-24b-instruct",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"mistralai\/mistral-small-3.1-24b-instruct",
     "bcp_47":"yue",
@@ -112167,6 +120483,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ar",
@@ -112279,6 +120602,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"as",
@@ -112314,6 +120644,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"awa",
@@ -112349,6 +120686,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"az",
@@ -112384,6 +120728,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"be",
@@ -112419,6 +120770,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"bho",
@@ -112573,6 +120931,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ceb",
@@ -112608,6 +120973,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ckb",
@@ -112888,6 +121260,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"fa",
@@ -113056,6 +121435,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"gu",
@@ -113140,6 +121526,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"hi",
@@ -113294,6 +121687,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"id",
@@ -113385,6 +121785,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ilo",
@@ -113518,6 +121925,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"jv",
@@ -113588,6 +122002,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"kk",
@@ -113623,6 +122044,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"km",
@@ -113658,6 +122086,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"kn",
@@ -113693,6 +122128,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ko",
@@ -113805,6 +122247,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"mai",
@@ -113840,6 +122289,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"mg",
@@ -113882,6 +122338,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ml",
@@ -113917,6 +122380,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"mr",
@@ -113952,6 +122422,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ms",
@@ -113994,6 +122471,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"my",
@@ -114029,6 +122513,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ne",
@@ -114120,6 +122611,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ny",
@@ -114211,6 +122709,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"or",
@@ -114246,6 +122751,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"pa",
@@ -114323,6 +122835,13 @@
     "metric":"chrf",
     "score":0.5674744623
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"pt",
@@ -114330,6 +122849,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"pt",
@@ -114365,6 +122891,13 @@
     "metric":"chrf",
     "score":0.6526848356
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ro",
@@ -114519,6 +123052,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"sd",
@@ -114554,6 +123094,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"si",
@@ -114645,6 +123192,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"so",
@@ -114687,6 +123241,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"sr",
@@ -114729,6 +123290,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"su",
@@ -114862,6 +123430,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ta",
@@ -115023,6 +123598,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ti",
@@ -115058,6 +123640,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"tr",
@@ -115100,6 +123689,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"uk",
@@ -115177,6 +123773,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"ur",
@@ -115212,6 +123815,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"uz",
@@ -115247,6 +123857,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":1.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"vi",
@@ -115471,6 +124088,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-mini",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"openai\/gpt-4.1-mini",
     "bcp_47":"yue",
@@ -115807,6 +124431,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ar",
@@ -115919,6 +124550,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"as",
@@ -115954,6 +124592,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"awa",
@@ -115989,6 +124634,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"az",
@@ -116024,6 +124676,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"be",
@@ -116059,6 +124718,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"bho",
@@ -116213,6 +124879,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ceb",
@@ -116248,6 +124921,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ckb",
@@ -116528,6 +125208,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"fa",
@@ -116696,6 +125383,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"gu",
@@ -116780,6 +125474,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"hi",
@@ -116934,6 +125635,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"id",
@@ -117025,6 +125733,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ilo",
@@ -117158,6 +125873,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"jv",
@@ -117228,6 +125950,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"kk",
@@ -117263,6 +125992,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"km",
@@ -117298,6 +126034,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"kn",
@@ -117333,6 +126076,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ko",
@@ -117445,6 +126195,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"mai",
@@ -117480,6 +126237,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"mg",
@@ -117522,6 +126286,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ml",
@@ -117557,6 +126328,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"mr",
@@ -117592,6 +126370,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ms",
@@ -117634,6 +126419,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"my",
@@ -117669,6 +126461,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ne",
@@ -117760,6 +126559,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ny",
@@ -117851,6 +126657,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"or",
@@ -117886,6 +126699,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"pa",
@@ -117963,6 +126783,13 @@
     "metric":"chrf",
     "score":0.5592753275
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"pt",
@@ -117970,6 +126797,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"pt",
@@ -118005,6 +126839,13 @@
     "metric":"chrf",
     "score":0.6189446172
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ro",
@@ -118159,6 +127000,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"sd",
@@ -118194,6 +127042,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"si",
@@ -118285,6 +127140,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"so",
@@ -118327,6 +127189,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"sr",
@@ -118369,6 +127238,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"su",
@@ -118502,6 +127378,13 @@
     "metric":"accuracy",
     "score":0.7
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ta",
@@ -118663,6 +127546,13 @@
     "metric":"accuracy",
     "score":0.6
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ti",
@@ -118698,6 +127588,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"tr",
@@ -118740,6 +127637,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"uk",
@@ -118817,6 +127721,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"ur",
@@ -118852,6 +127763,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"uz",
@@ -118887,6 +127805,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"vi",
@@ -119111,6 +128036,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4.1-nano",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"openai\/gpt-4.1-nano",
     "bcp_47":"yue",
@@ -119447,6 +128379,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ar",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ar",
@@ -119559,6 +128498,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"as",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"as",
@@ -119594,6 +128540,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"awa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"awa",
@@ -119629,6 +128582,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"az",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"az",
@@ -119664,6 +128624,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"be",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"be",
@@ -119699,6 +128666,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"bho",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"bho",
@@ -119853,6 +128827,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ceb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ceb",
@@ -119888,6 +128869,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ckb",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ckb",
@@ -120168,6 +129156,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"fa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"fa",
@@ -120336,6 +129331,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"gu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"gu",
@@ -120420,6 +129422,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"hi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"hi",
@@ -120574,6 +129583,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"id",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"id",
@@ -120665,6 +129681,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ilo",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.3
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ilo",
@@ -120798,6 +129821,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"jv",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"jv",
@@ -120868,6 +129898,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"kk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"kk",
@@ -120903,6 +129940,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"km",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"km",
@@ -120938,6 +129982,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"kn",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"kn",
@@ -120973,6 +130024,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ko",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ko",
@@ -121085,6 +130143,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"mai",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"mai",
@@ -121120,6 +130185,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"mg",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"mg",
@@ -121162,6 +130234,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ml",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ml",
@@ -121197,6 +130276,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"mr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"mr",
@@ -121232,6 +130318,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ms",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ms",
@@ -121274,6 +130367,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"my",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"my",
@@ -121309,6 +130409,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ne",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ne",
@@ -121400,6 +130507,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ny",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ny",
@@ -121491,6 +130605,13 @@
     "metric":"accuracy",
     "score":1.0
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"or",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.5
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"or",
@@ -121526,6 +130647,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"pa",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"pa",
@@ -121603,6 +130731,13 @@
     "metric":"chrf",
     "score":0.535772663
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ps",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"pt",
@@ -121610,6 +130745,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"pt",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"pt",
@@ -121645,6 +130787,13 @@
     "metric":"chrf",
     "score":0.6773057972
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"qu",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.2
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ro",
@@ -121799,6 +130948,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"sd",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"sd",
@@ -121834,6 +130990,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"si",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"si",
@@ -121925,6 +131088,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"so",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.4
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"so",
@@ -121967,6 +131137,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"sr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"sr",
@@ -122009,6 +131186,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"su",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"su",
@@ -122142,6 +131326,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ta",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.8
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ta",
@@ -122303,6 +131494,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ti",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.0
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ti",
@@ -122338,6 +131536,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"tr",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"tr",
@@ -122380,6 +131585,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"uk",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.9
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"uk",
@@ -122457,6 +131669,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"ur",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.7
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"ur",
@@ -122492,6 +131711,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"uz",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"uz",
@@ -122527,6 +131753,13 @@
     "metric":"accuracy",
     "score":0.9
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"vi",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.6
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"vi",
@@ -122751,6 +131984,13 @@
     "metric":"accuracy",
     "score":0.8
   },
+  {
+    "model":"openai\/gpt-4o-mini",
+    "bcp_47":"yue",
+    "task":"mgsm",
+    "metric":"accuracy",
+    "score":0.1
+  },
   {
     "model":"openai\/gpt-4o-mini",
     "bcp_47":"yue",