Update README.md

README.md

@@ -60,13 +60,16 @@ tags:
 ---
 
 This monolingual (English) NLI dataset is designed for performing Natural Language Inference, and is particularly Fact-Checking oriented.
+Dev split is oriented to teach the model how to deal well with pure NLI (ANLI is well designed for this task) and test his general knwoledge with VitaminC, which is known for its robustness for this task.
 
 It contains:
-- 8 whole datasets in the training split (SNLI, MNLI, FEVER, QNLI, WNLI, SciTail, RTE, VitaminC);
-- 120 examples for the dev split:
-  - 60 from the Climate-FEVER dataset in order to test zero-shot knowledge inference in a specific domain;
-  - 60 from the ANLI training splits in order to test pure NLI skills of the model.
-- 2 whole datasets in the test split (3 training splits of the ANLI dataset and the whole Climate-FEVER dataset).
+- 14.5k examples for the dev split of which:
+  - 848 from ANLI train_r1;
+  - 2273 from ANLI train_r2;
+  - 5023 from ANLI train_r3;
+  - 6306 from VitaminC dev.
+- 4 test splits (the 3 test splits of the ANLI dataset and 10% of the VitaminC test split).
+- The remaining data composes the train split.
 
 Datasets references:
 - SNLI: https://huggingface.co/datasets/stanfordnlp/snli