Classify Text (SetFit)

@huggingface1 VersionClassify

Few-shot SetFit classifier that scores a String against a class set fitted from a handful of examples, emitting ranked [Classification], a single Classification, or [Double] in sorted class order. For few examples per class with no LLM endpoint.

How it fits

Typical backends

Few-shot intent tagging over a text file, published as JSON.

Topic routing on extracted document text, gated by a branch.

Caveats

State lifetimeThe head is fitted once at startup over the body bound at model and is fixed for the deployment; changing the example set or backbone requires a restart, and only top_k is read per message.
Hard constraintThe startup fit needs paired example_texts and example_labels of equal length with at least two distinct labels, or it fails fast at startup.
I/O contractOne ranked result is produced per text and every arm projects it, so the arms agree class-for-class; on the [Double] arm scores align to the sorted class set order, which downstream must pin.
AccuracyEight to sixteen balanced examples per class is usually enough; the contrastive fit is sensitive to per-class skew.
FallbackEmpty or whitespace-only String input returns the empty form of the resolved arm (an empty ranked list, a zero-confidence empty-label Classification, or an all-zero score vector) without running the model.

Versions

d4e4f498latestdefault
Automated release
6/25/2026