| ====== Perplexity statistics ====== | |
| Mean PPL(Q) : 17.594278 ± 0.144215 | |
| Mean PPL(base) : 14.009216 ± 0.118474 | |
| Cor(ln(PPL(Q)), ln(PPL(base))): 91.45% | |
| Mean ln(PPL(Q)/PPL(base)) : 0.227858 ± 0.003452 | |
| Mean PPL(Q)/PPL(base) : 1.255907 ± 0.004335 | |
| Mean PPL(Q)-PPL(base) : 3.585061 ± 0.059859 | |
| ====== KL divergence statistics ====== | |
| Mean KLD: 0.553217 ± 0.001919 | |
| Maximum KLD: 16.471909 | |
| 99.9% KLD: 6.719028 | |
| 99.0% KLD: 3.491813 | |
| 99.0% KLD: 3.491813 | |
| Median KLD: 0.338121 | |
| 10.0% KLD: 0.015783 | |
| 5.0% KLD: 0.003790 | |
| 1.0% KLD: 0.000329 | |
| Minimum KLD: 0.000001 | |
| ====== Token probability statistics ====== | |
| Mean Δp: -6.554 ± 0.049 % | |
| Maximum Δp: 99.861% | |
| 99.9% Δp: 72.278% | |
| 99.0% Δp: 38.259% | |
| 95.0% Δp: 14.466% | |
| 90.0% Δp: 6.153% | |
| 75.0% Δp: 0.180% | |
| Median Δp: -0.650% | |
| 25.0% Δp: -10.364% | |
| 10.0% Δp: -29.955% | |
| 5.0% Δp: -44.578% | |
| 1.0% Δp: -74.731% | |
| 0.1% Δp: -95.812% | |
| Minimum Δp: -99.968% | |
| RMS Δp : 19.812 ± 0.070 % | |
| Same top p: 67.032 ± 0.124 % | |