Update README.md

README.md

@@ -1,30 +1,37 @@
----
-license: mit
-configs:
-- config_name: default
-  data_files:
-  - split: train
-    path: data/train-*
-dataset_info:
-  features:
-  - name: query_id
-    dtype: int64
-  - name: text
-    dtype: string
-  - name: document_ids
-    sequence: string
-  - name: scores
-    sequence: float64
-  - name: means
-    dtype: float64
-  - name: stds
-    dtype: float64
-  - name: maxmins
-    dtype: float64
-  splits:
-  - name: train
-    num_bytes: 351849640
-    num_examples: 500000
-  download_size: 198040061
-  dataset_size: 351849640
----
+---
+license: mit
+configs:
+- config_name: default
+  data_files:
+  - split: train
+    path: data/train-*
+dataset_info:
+  features:
+  - name: query_id
+    dtype: int64
+  - name: text
+    dtype: string
+  - name: document_ids
+    sequence: string
+  - name: scores
+    sequence: float64
+  - name: means
+    dtype: float64
+  - name: stds
+    dtype: float64
+  - name: maxmins
+    dtype: float64
+  splits:
+  - name: train
+    num_bytes: 351849640
+    num_examples: 500000
+  download_size: 198040061
+  dataset_size: 351849640
+---
+
+This dataset is based on the translated MMARCO dataset. 
+500k queries (from [queries](https://huggingface.co/datasets/akhooli/ar_mmarco_queries)) were selected and scored against 32 documents each 
+(based on English scores) from 
+[Ar MMARCO Docs](https://huggingface.co/datasets/akhooli/ar_mmarco_docs) (translated) using JinaAI multilingual reranker.  
+The dataset can be used to train knowledge distillation ColBERT models and some are available under the models section.  
+Note: translation quality is not good enough and scores are not spread apart. Some statistics were applied to curate a subset with wide distribution but this is not enough.