Recognize Named Entities NER (HuggingFace)

@huggingface1 VersionExtract

Runs zero-shot GLiNER named-entity recognition on each String and emits the spans that clear a confidence floor as either anonymous record spans or [NamedEntity] carriers, depending on the downstream connection.

How it fits

Typical backends

Document entity extraction. Read text from a file, extract the configured entity types, and publish them over HTTP.

OCR-to-entities. Extract text from scanned images and mine named entities from the recognised text.

Transcript entity mining. Transcribe audio and extract entities from each transcript line.

Caveats

AccuracyThe entity vocabulary is an open list of natural-language types matched zero-shot, so an empty list emits nothing and the exact wording of a label shifts recall.
I/O contractEach span carries character offsets into the input string and a type that is the matched vocabulary entry rather than a fixed schema; the entity output form follows whatever the downstream connection requires.
Parameter interactionThe entity vocabulary is re-read on every call, while the confidence floor, the model selection, and the inference device are captured once at startup.
CompatibilityThe default checkpoint targets general multi-task accuracy; in-domain fine-tunes for biomedical, news, PII, or multilingual text do better on their domain, and the model also runs acceptably on CPU.

Versions

dd2a7a1edefaultlatest
Automated release
5/12/2026