Update README.md

README.md

@@ -40,7 +40,7 @@ This is the Triplit Critic model presented in the ACL 2023 paper [RED^{FM}: a Fi
         url = "https://arxiv.org/abs/2306.09802",
     }
 
-The Triplit Critic is based on mdeberta-v3-base and it was trained as a multitask system to filter triplets as well as on the XNLI dataset. The model weights contain the two classification heads, however loading it using the huggingface library will only load those for Triplet filtering (ie. a binary classification head), if one wants to use it for XNLI it needs a custom script. While it is defined and trained as a classification system, we use the positive score (ie. Label_1) as the confidence score for a triplet. For SRED<sup>FM</sup> the confidence score thresshold was set at 0.75.
+The Triplit Critic is based on mdeberta-v3-base and it was trained as a multitask system to filter triplets (triplet filtering data used can be found in [Triplet-Critic-Data](https://huggingface.co/Babelscape/mdeberta-v3-base-triplet-critic-xnli/tree/main/Triplet-Critic-Data)) as well as on the XNLI dataset. The model weights contain the two classification heads, however loading it using the huggingface library will only load those for Triplet filtering (ie. a binary classification head), if one wants to use it for XNLI it needs a custom script. While it is defined and trained as a classification system, we use the positive score (ie. Label_1) as the confidence score for a triplet. For SRED<sup>FM</sup> the confidence score thresshold was set at 0.75.