From the model card
Fields below are copied from the tags and counters on the HuggingFace repository facebook/bart-large-mnli at our last fetch. They are set by the uploader, not verified by us; rows with no tag are omitted. How this page is made.
- Publisher (HF namespace)
- Pipeline tag
- zero-shot-classification
- Library
- Transformers
- Framework tags
- PyTorch, JAX, Rust (candle)
- Weight formats
- safetensors
- License tag
mit— read the license file in the repo before relying on it- Papers cited
- arXiv:1910.13461, arXiv:1909.00161
- Datasets declared
- multi_nli
- Downloads (HF counter at last fetch)
- 3,078,289
- Likes (HF counter at last fetch)
- 1,605
- Model card
- https://huggingface.co/facebook/bart-large-mnli
Use cases
- Zero-shot topic classification across arbitrary label sets
- Intent detection when labeled data is unavailable
- Content tagging pipelines that need to adapt to new categories quickly
- Filtering text by theme without training a custom classifier
Pros
- No training data required — works with any label defined in natural language
- BART architecture handles longer texts better than BERT-class models
- MIT licensed
- Widely benchmarked zero-shot baseline in the NLP literature
Cons
- Slower than trained classifiers — runs NLI inference per label
- Quality degrades when labels are ambiguous or abstractly defined
- Doesn't scale to hundreds of candidate classes without multi-label optimizations
- DeBERTa-v3 and Mistral-based zero-shot classifiers now outperform it
Tags
transformerspytorchjaxrustsafetensorsbarttext-classificationzero-shot-classificationdataset:multi_nliarxiv:1910.13461arxiv:1909.00161license:mitendpoints_compatibleregion:usdeploy:sagemakerdeploy:azure