meddies-pii-v2-onnx

Meddies PII v2 ONNX

<p align="left"> <a href="https://huggingface.co/spaces/Meddies/meddies-pii-extractor"><img alt="Try the browser demo" src="https://img.shields.io/badge/Demo-0F172A?style=flat&logo=googlechrome&logoColor=white"></a> <a href="https://huggingface.co/datasets/Meddies/meddies-pii-v2"><img alt="Training dataset" src="https://img.shields.io/badge/Dataset-024C6F?style=flat&logo=huggingface&logoColor=white"></a> <a href="https://huggingface.co/Meddies/meddies-pii-v2"><img alt="PyTorch model" src="https://img.shields.io/badge/PyTorch-1ABBC7?style=flat&logo=huggingface&logoColor=white"></a> <a href="https://creativecommons.org/licenses/by-nc/4.0/"><img alt="CC BY-NC 4.0" src="https://img.shields.io/badge/license-CC--BY--NC--4.0-475569"></a> </p>

The ONNX build of Meddies PII v2. It finds nine types of personally identifiable information across 17 languages and returns exact character spans for review and redaction.

[!NOTE] ONNX companion

This repository mirrors the safety, license, hero, and evaluation summary from the main release. Use the main card for training lineage and the full benchmark tables.

[!CAUTION] Keep a human in the loop

This model can miss identifiers. It cannot certify that text is safe to share. Test it on your documents and keep human review in any regulated workflow.

[!NOTE] Commercial licensing

CC BY-NC 4.0 applies to the rights MeddiesAI can license in this release. Contact [email protected] about commercial terms for those contributions. The underlying LiquidAI materials remain subject to LFM Open License v1.0.

<p align="center"> <img src="./assets/hero-meddies-pii-v2.png" alt="Meddies PII v2 detects nine PII entity families across 17 languages and returns exact character spans" width="100%" /> </p>

Why this build

Clinical text often needs to stay on a local server or inside the user's browser. This build runs the Meddies PII v2 checkpoint through ONNX Runtime while keeping the same tokenizer, label map, and BIOES decoder as the main release.

What it ships

Pass in raw text. The model returns typed character spans for address, company_name, date, email_address, human_name, id_number, phone_number, private_url, and secret.

• Full precision graph with model.onnx, which preserved parity with the PyTorch adapter on the 200-row check.
• Current int8 graph with model.int8.onnx, which is smaller but changed decoded spans on 12 of 200 parity rows.
• Shared decoding path with the tokenizer, offset mapping, label map, and Viterbi BIOES decoder included in the release.
• Local execution through the browser demo, which processes raw clinical text on the user's device.

Quick start

import sys

from huggingface_hub import snapshot_download

path = snapshot_download("Meddies/meddies-pii-v2-onnx")
sys.path.insert(0, path)

from usage_onnxruntime import extract

spans = extract(
    "Bệnh nhân Nguyễn Văn A, SĐT 0912 345 678, tái khám ngày 12/08/2026."
)

for span in spans:
    print(span.label, span.start, span.end, span.text)

Choose the ONNX graph

Both graphs take dynamic int64 input IDs and attention masks. They return logits for 37 BIOES classes. The shipped helper loads the full precision graph by default.

[!WARNING] Browser memory

The 647 MB int8 graph is still too large for some memory-constrained browsers. Treat it as an ONNX release artifact, not a universal browser default. A smaller graph needs a fresh quality check before it replaces this one.

The browser demo runs on the user's device. Raw clinical text never leaves the device.

Evaluation and parity

The benchmark numbers below describe the main Meddies PII v2 checkpoint. The full precision graph matched that checkpoint on all 200 parity rows. The current int8 graph changed decoded spans on 12 of 200 rows. Parity checks measure span agreement, not accuracy on a new dataset.

<p align="center"> <img src="./assets/evaluation-meddies-pii-v2-onnx.png" alt="Meddies PII v2 checkpoint quality with full precision and int8 ONNX parity shown separately" width="100%" /> </p>

Exact typed F1 requires the label and both character boundaries to match. Containment typed F1 accepts a correct label when one span contains the other. Read the main model card for the full evaluation cell and entity family tables.

How it was built

The graphs export the Meddies PII v2 checkpoint with dynamic sequence axes and the same tokenizer and 37-tag BIOES map. The main release owns the model lineage, training data, checkpoint selection, and benchmark protocol.

Where it fits

• Local ONNX Runtime services that need typed spans instead of generated text.
• Browser review tools where raw text must stay on the user's device.
• Deployment experiments that measure graph size, speed, memory, and span parity on the real target.

Limits

• The int8 graph changed decoded spans on 12 of 200 parity rows.
• The 647 MB int8 graph can exceed the memory budget of constrained browsers.
• Neither graph processes images or document layout. Run optical character recognition first for scans.
• WebGPU support and speed depend on the browser, device, and ONNX Runtime Web build.
• The model can still miss unsupported identifiers or inputs designed to evade detection.

License

MeddiesAI applies CC BY-NC 4.0 to the copyright and related rights it can license in this release. See LICENSE.

The underlying LiquidAI materials remain under LFM Open License v1.0. Upstream attribution and the MeddiesAI change record are preserved in NOTICE. Review both licenses before redistribution.

Feedback

Send runtime failures, unsupported operators, WebGPU incompatibilities, memory crashes, or span differences from the PyTorch model. Include the graph filename, ONNX Runtime version, execution provider, device, and a safe synthetic example.

Open a thread in the Community tab or email [email protected]. Never post real patient text or credentials.

Collaboration

We are looking for teams that can test ONNX Runtime and WebGPU on real hospital hardware without exposing patient data. Email [email protected] to work with us.

Citation

@misc{meddies-pii-v2-onnx,
  title={Meddies PII v2 ONNX for multilingual clinical de-identification},
  author={MeddiesAI},
  year={2026},
  url={https://huggingface.co/Meddies/meddies-pii-v2-onnx}
}