README.md

---
datasets:
- multi_nli
- snli
- scitail
language:
- en
metrics:
- accuracy
- f1
pipeline_tag: zero-shot-classification
---
# RoBERTa NLI (Natural Language Inference) 
This model is a fine-tuned model of [roberta-large](https://huggingface.co/roberta-large) after being trained on a **mixture of NLI datasets**.

This model can classify a pair of sentence (a <u>premise</u> and a <u>claim</u>) into 3 classes:
- 'entailment': the claim can logically be inferred from the premise
- 'contradiction': the claim contradicts the premise
- 'neutral': the premise is unrelated or do not provide sufficient information to validate the claim

This model can also be used for **zero-shot classification tasks** !
Please take a look at this [repo](https://github.com/AntoineBlanot/zero-nlp) for more information on zero-shot classification tasks.

# Usage
This model has been trained in an efficient way and thus cannot be load directly from HuggingFace's hub. To use that model, please follow instructions on this [repo](https://github.com/AntoineBlanot/efficient-llm).

For **zero-shot classification** tasks, please take a look at this [repo](https://github.com/AntoineBlanot/zero-nlp).

# Data used for training
- multi_nli
- snli
- scitail

# Evaluation results

| Data | Accuracy |
|:---:|:---------:|
| MNLI (val. m) | 0.894 |
| MNLI (val. mm) | 0.895 |
| SNLI (val.) | 0.920 |
| SciTail (val.) | 0.934 |