Deberta V3 Large Mnli

🚀 DeBERTa-v3 (large) fine-tuned to Multi-NLI (MNLI)

This model is designed for Textual Entailment (NLI), aiming to predict whether textA is supported by textB. Specifically, it performs a 2-way classification where the relationship between textA and textB can be entail, neutral, contradict.

🚀 Quick Start

Model Input and Output

• Input: (textA, textB)
• Output: prob(entail), prob(contradict)

Note that during training, all 3 labels (entail, neutral, contradict) were used. However, for this model, the neutral output head has been removed.

✨ Features

Model Details

💻 Usage Examples

Basic Usage

from transformers import AutoTokenizer, AutoModelForSequenceClassification
tokenizer = AutoTokenizer.from_pretrained("potsawee/deberta-v3-large-mnli")
model = AutoModelForSequenceClassification.from_pretrained("potsawee/deberta-v3-large-mnli")

textA = "Kyle Walker has a personal issue"
textB = "Kyle Walker will remain Manchester City captain following reports about his private life, says boss Pep Guardiola."

inputs = tokenizer.batch_encode_plus(
    batch_text_or_text_pairs=[(textA, textB)],
    add_special_tokens=True, return_tensors="pt",
)
logits = model(**inputs).logits # neutral is already removed
probs = torch.softmax(logits, dim=-1)[0]
# probs = [0.7080, 0.2920], meaning that prob(entail) = 0.708, prob(contradict) = 0.292

📄 License

This project is licensed under the Apache-2.0 license.

📚 Documentation

Citation

@article{manakul2023selfcheckgpt,
  title={Selfcheckgpt: Zero-resource black-box hallucination detection for generative large language models},
  author={Manakul, Potsawee and Liusie, Adian and Gales, Mark JF},
  journal={arXiv preprint arXiv:2303.08896},
  year={2023}
}