Wismut/openmed-onnx

OpenMed PII — ONNX

ONNX conversions of OpenMed PII token-classification models, packaged for the nym anonymization CLI's OpenMed NER backend.

These are not new models — they are straight ONNX exports of OpenMed's DeBERTa-v2 PII checkpoints (106-label BIO PII taxonomy), provided because OpenMed publishes PyTorch safetensors but no ONNX. All credit for the models belongs to the OpenMed project; see Attribution below.

Contents

Each subfolder is a self-contained model (config.json, tokenizer.json, model.onnx, …):

Subfolder	Source model	Params	Files
small/	OpenMed/OpenMed-PII-SuperClinical-Small-44M-v1	44M	model.onnx (fp32) + model_int8.onnx (dynamic int8)
base/	OpenMed/OpenMed-PII-SuperClinical-Base-184M-v1	184M	model.onnx (fp32)
large/	OpenMed/OpenMed-PII-SuperClinical-Large-434M-v1	434M	model.onnx (fp32)

int8 note: dynamic int8 quantization is reliable only for the small model. It collapses the base/large models (every token predicted as one label), so those ship fp32 only. See the nym docs for details.

Conversion

Exported with 🤗 optimum-onnx (optimum-cli export onnx --task token-classification --opset 17); int8 via onnxruntime dynamic quantization. Reproducible with scripts/convert_openmed_onnx.sh in the nym repo.

Usage with nym

[ner]
enabled = true
backend = "openmed"
openmed_model = "Wismut/openmed-onnx/small"   # or /base, /large
threshold = 0.5

nym downloads and caches the selected subfolder automatically on first run.

Inputs / outputs

• Inputs: input_ids, attention_mask (int64, [batch, seq]); max sequence 512.
• Output: logits [batch, seq, 106]. Decode with the id2label map in config.json using BIO scheme.

Attribution

These artifacts are derived from OpenMed by Maziyar Panahi, released under the Apache-2.0 License. Original project: https://github.com/maziyarpanahi/openmed.

@misc{panahi2025openmedneropensourcedomainadapted,
      title={OpenMed NER: Open-Source, Domain-Adapted State-of-the-Art Transformers for Biomedical NER Across 12 Public Datasets},
      author={Maziyar Panahi},
      year={2025},
      eprint={2508.01630},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2508.01630},
}

License

Apache-2.0, inherited from the upstream OpenMed models. See the LICENSE file in this repo. The conversion adds no additional restrictions.