| --- |
| license: apache-2.0 |
| library_name: gliner2 |
| base_model: fastino/gliner2-privacy-filter-PII-multi |
| datasets: |
| - synthetic |
| language: |
| - en |
| - fr |
| - es |
| - de |
| - it |
| - pt |
| - nl |
| pipeline_tag: token-classification |
| tags: |
| - onnx |
| - pii |
| - ner |
| - privacy |
| - redaction |
| - gliner |
| - gliner2 |
| - information-extraction |
| - span-extraction |
| --- |
| |
| # GLiNER2-PII ONNX |
|
|
| This model is meant for use within the Redacta chrome extension, and this ONNX export was specifically meant for browser use, removing dependencies and preserving behavior. |
|
|
| GLINER2-PII ONNX is 2.75x smaller than the original GLINER2-PII with nearing performance. |
|
|
| ## Benchmark Results (SPY) |
|
|
| Evaluated on the [SPY benchmark](https://aclanthology.org/2025.naacl-srw.23/) (Savkin et al., 2025) with exact-match span-level metrics: |
|
|
| | Model | Legal P | Legal R | Legal F1 | Medical P | Medical R | Medical F1 | **Avg F1** | |
| |---|:---:|:---:|:---:|:---:|:---:|:---:|:---:| |
| | **yethdev/gliner2-pii-onnx** | .383 | **.861** | **.530** | .396 | **.846** | **.539** | **.535** | |
| | fastino/gliner2-pii-v1 | .346 | .750 | .473 | .369 | .686 | .480 | .477 | |
| | nvidia/gliner-PII | .343 | .452 | .390 | .368 | .465 | .411 | .400 | |
| | urchade/gliner\_multi\_pii-v1 | **.467** | .317 | .377 | **.518** | .351 | .419 | .398 | |
| | openai/privacy-filter | .242 | .656 | .354 | .287 | .692 | .406 | .380 | |
|
|
| Setup for the yethdev/gliner2-pii-onnx row: full splits (4,197 legal questions, 4,491 medical consultations), entities materialized with `faker_random_seed=42`, micro-averaged exact character-level span match on the model's raw span output (no text-level deduplication). The model was queried with the labels `email`, `id number`, `name`, `phone number`, `address`, `url`, `username` mapped 1:1 to SPY categories, at confidence threshold 0.3; label phrasing and threshold were selected on a held-out 5% development sample. Inference ran the released `model.quantized.onnx` via onnxruntime (CPU), which matches the PyTorch engine's span sets on text up to 2,900 subwords with max confidence drift under 0.03. |
|
|
| ## License & attribution |
|
|
| Apache-2.0 - All model weights are derived from |
| [fastino/gliner2-privacy-filter-PII-multi](https://huggingface.co/fastino/gliner2-privacy-filter-PII-multi) |
| by Fastino Labs. |
