README.md

---
title: GLiNER2.5-multi-Decide — Multilingual Text Decisions
emoji: 🌍
colorFrom: red
colorTo: blue
sdk: gradio
sdk_version: 6.28.0
app_file: app.py
short_description: Multilingual zero-shot text classification, any label set
python_version: "3.12"
startup_duration_timeout: 30m
---

Multilingual operational text classification with
[fastino/GLiNER2.5-multi-Decide](https://huggingface.co/fastino/GLiNER2.5-multi-Decide).

One 287M-parameter multilingual checkpoint (mDeBERTa-v3-base) answers **any label set you pass
at call time** — intent, routing, sentiment, priority, policy, multi-label tags — in a single
forward pass. No prompt template, no generated tokens. Use it when the text is *not English*:
the same call scores several decision heads at once, and single-label heads return one string
while multi-label heads return every label above their threshold.

## How to use it

1. Paste the text to decide about — any language.
2. Edit the **decisions table** — one row per decision head:
   - **Decision** — the head's name (`intent`, `sentiment`, `urgency`, …).
   - **Labels** — comma-separated candidate labels. Use `"0", "1", …` strings for an ordinal scale.
   - **Multi-label** — tick when several labels can apply at once (review aspects, topics).
   - **Threshold** — the confidence cutoff for multi-label heads.
3. Press **Decide**. Every head is scored together in the same call.

**Advanced → Per-decision overrides** takes a JSON object merged into a head by name, for a
question over the passage (`prompt`), label descriptions, or other per-head config:

```json
{"answer": {"prompt": "Did the treaty enter into force in 1992?"}}
```

```json
{"intent": {"labels": {"card_pin_change": "The customer wants a new PIN", "card_lost": "The physical card is missing"}}}
```

## About the model

GLiNER2.5-multi-Decide is the multilingual specialist of the GLiNER2.5 family for operational
decisions: customer and banking intent, travel and clinic requests, review sentiment, document
type, email and ticket routing, human handoff, moderation, severity, urgency, and spam. It
scores 56.7% exact-match on the 17-domain held-out
[`fastino/fast-decisions`](https://huggingface.co/datasets/fastino/fast-decisions) suite, and is
loaded through the `gliner2` library's `AutoExtractor`.

It is **not** a general-purpose model: it does not reason, explain, or answer open questions.

## Links

- Model: [fastino/GLiNER2.5-multi-Decide](https://huggingface.co/fastino/GLiNER2.5-multi-Decide)
- Code: [github.com/fastino-ai/GLiNER2](https://github.com/fastino-ai/GLiNER2)
- Paper: [GLiNER2 (arXiv:2507.18546)](https://arxiv.org/abs/2507.18546)

## License

Apache-2.0 (model). The example texts are the ones published on the model card.