This is a Deberta-v3 large NLI model that has been finetuned on augmented ANLI data to improve performance for faithfulness evaluation, particularly in a knowledge-grounded dialog setting. See https://aclanthology.org/2023.acl-short.79 for details. The model itself is based on https://huggingface.co/MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli For a convenient interface for scoring see our repository: https://github.com/julmaxi/with_a_little_push
Note that this is not the original model used in the paper but a retrained variant that has slightly better average performance.
- Downloads last month
- 48