README.md

---
language:
- en

pipeline_tag: text-classification
---

# Span NLI BERT (base)

This is a **BERT-base** model ([`bert-base-uncased`][2]) fine-tuned on the [**ContractNLI**][3] dataset (non-disclosure agreements) with the **Span NLI BERT** model architecture, 
from [*ContractNLI: A Dataset for Document-level Natural Language Inference for Contracts* (Koreeda and Manning, 2021)][1].

For a hypothesis, the **Span NLI BERT** model predicts NLI labels and identifies evidence for documents as premises. 
Spans of documents should be pre-annotated; evidence is always full sentences or items in an enumerated list in the document. 

For details of the architecture and usage of the relevant training/testing scripts, check out the paper and their [Github repo][4]. 
This model is fine-tuned according to the hyperparameters in `data/conf_base.yml` in their repo, 
which differs from their hyperparameters that produced the best dev scores as noted in the Appendix of the paper. 

ArXiv: <https://arxiv.org/abs/2110.01799>

[1]: https://aclanthology.org/2021.findings-emnlp.164/
[2]: https://huggingface.co/bert-base-uncased
[3]: https://stanfordnlp.github.io/contract-nli/
[4]: https://github.com/stanfordnlp/contract-nli-bert