---
license: mit
base_model: urchade/gliner_multi-v2.1
pipeline_tag: token-classification
library_name: onnx
language:
- multilingual
tags:
- ner
- gliner
- onnx
---

# ONNX - gliner_multi-v2.1
FP32 = model.onnx

INT8 = model_quantized.onnx

## Usage

```
# onnxruntime-gpu for run it on GPU
pip install onnxruntime sentencepiece numpy protobuf
```

- [src/helper.py](src/helper.py) = onnx session builder (provider selection, threads, memory options)
- [src/example.py](src/example.py) = full pipeline: word split, sentencepiece tokenization, span enumeration, inference, greedy non-overlap selection

```
python3 src/example.py
```

```
0.970778 | person | Linus Torvalds
0.603953 | organization | Linux
0.978727 | place | Helsinki
0.859634 | organization | Linux Foundation
0.944801 | person | 田中太郎
0.674417 | organization | 東京大学
0.953844 | place | Tokyo
```

Note:
- No transformers/tokenizers dependency: the tokenizer is `spm.model` loaded via proto with the patch<br>
`normalizer_spec.add_dummy_prefix = False` (required for the word-level alignment).
- Entity types are free text, passed in the input sequence:<br>
`[cls] [ent] type... [ent] type... [sep] word... [eos]`, with `[ent]` = 250103, `[sep]` = 250104.
- `words_mask` marks the first subword of each word with the word number (1-based), 0 elsewhere.
- Digits are tokenized one at a time (`\d|\D+`), CJK characters are split one word each.
- Spans are enumerated per word up to width 12, score = sigmoid(logit), greedy selection without overlap.
- Normalize input with NFKC for consistent multilingual scores.