MultiNli — Inference Multi-genre
Massive open-source corpus to train models to detect the nicking, neutrality, or contradiction between two sentences in different textual genres.
Description
MultiNli (Multi-Genre Natural Language Inference) is a dataset of more than 430,000 pairs of annotated sentences with logical inference relationships: nicking, neutral, contradiction. It covers a wide variety of genres, ranging from government texts to fictions, which makes it possible to test the ability to generalize models between different styles of language.
What is this dataset for?
- Train models to detect logical inference between two sentences (NLI)
- Evaluate the robustness of NLP models in various fields
- Building QA or dialogue systems that are sensitive to semantic context
Can it be enriched or improved?
Yes. It is possible to complete this dataset with translations to create a multilingual version or to annotate it with confidence scores for each label. It is also possible to increase linguistic diversity by remixing genres or by adding automatically generated hypotheses.
🔎 In summary
🧠 Recommended for
- NLP researchers
- Bert-like models
- Cross-domain evaluation
🔧 Compatible tools
- Hugging Face Datasets
- PyTorch
- TensorFlow
- Scikit-learn
💡 Tip
Filter examples with a -1 label before training to avoid noisy data.
Frequently Asked Questions
What is the difference between MultiNLI and SNLI?
SNLI mostly covers image descriptions, while MultiNLI includes numerous text styles, making it more general and robust.
Can MultiNli be used to evaluate multilingual models?
Not directly, as it's only in English. But it can be used as a basis for multilingual adaptation through translation.
Is the dataset suitable for fine-tuning large models like GPT or BERT?
Yes, it is particularly useful for refining Transformer-type models for inference or semantic comprehension tasks.




