GLiNER multi v2.1 (ONNX)
A zero-shot entity detector: it returns the spans of a text that match a list of entity-type labels
supplied at call time. This repository holds the float32 ONNX graph of
urchade/gliner_multi-v2.1 at revision
443d26d6, exported for ONNX Runtime, with the configuration and tokenizer files beside it.
What it detects
Whatever types the caller supplies as labels. The model finds the spans of those labels and no type a label does not describe.
Languages
Multilingual. The labels are supplied by the caller in the language of the text; the measurement below is in French.
Measured recall
On a held-out split of 60 synthetic French clinical notes with 2 423 gold entity spans (no real patient data), prompting with a 16-label clinical list at span score 0.3, measured on 2026-11-03:
| Measure | Value |
|---|---|
| Overlap recall | 0.954 |
| Exact recall | 0.764 |
| Overlap precision | 0.565 |
The split is small and synthetic; the figures do not carry over to other domains or to longer, less regular text.
Intended use
Extracting entity spans before a text is de-identified or shared: person names, organisations, places, addresses, dates, telephone numbers, email addresses and identifiers, and any other type the caller names. A person checks the spans before the text goes anywhere; recall is below 100 %.
Files
| File | Content |
|---|---|
model.onnx |
the float32 graph (opset 19) |
gliner_config.json |
the model's settings |
tokenizer.json, tokenizer_config.json |
the tokenizer the graph runs with |
Licence
The graph is a derivative of urchade/gliner_multi-v2.1 (Apache-2.0) and of its encoder
microsoft/mdeberta-v3-base (MIT). It is
published under the Apache-2.0 licence (LICENSE); the encoder's MIT notice is in NOTICE. The
graph was modified: exported to ONNX (opset 19) from the PyTorch checkpoint.
References
- Zaratiana et al., GLiNER: Generalist Model for Named Entity Recognition using Bidirectional Transformer, arXiv:2311.08526
urchade/gliner_multi-v2.1