OpenMed / privacy-filter-multilingual

privacy-filter-multilingual

Fine-tuned openai/privacy-filter for fine-grained PII extraction across 54 categories in 16 languages .

• Base model : openai/privacy-filter — 1.4B-parameter MoE (50M active per token), BIOES token-classification head
• Task : Token classification for PII detection (BIOES scheme)
• Languages (16) : Arabic, Bengali, Chinese, Dutch, English, French, German, Hindi, Italian, Japanese, Korean, Portuguese, Spanish, Telugu, Turkish, Vietnamese
• Training data : Multilingual mix from AI4Privacy — pii-masking-200k , pii-masking-400k , and open-pii-masking-500k-ai4privacy , language-balanced
• Recipe : opf train (OpenAI's official fine-tuning CLI) — full fine-tune, AdamW, balanced language sampling, 5 epochs, bf16
• Labels : 54 PII categories → 217 BIOES classes (1 O + 54 × B/I/E/S)

The base model ships with 8 coarse PII categories and English-only training. This model trades that for a 6.75× more granular vocabulary spanning identity, contact, address, financial, vehicle, digital, and crypto labels — all evaluated across 16 languages.

Family at a glance. Same architecture, three runtimes:

• PyTorch (this repo) — CPU + CUDA, anywhere transformers runs.
• MLX BF16 — OpenMed/privacy-filter-multilingual-mlx — Apple Silicon, full precision.
• MLX 8-bit — OpenMed/privacy-filter-multilingual-mlx-8bit — Apple Silicon, smaller + faster.

Quick start

With OpenMed — recommended

OpenMed gives you extract_pii() / deidentify() with built-in BIOES Viterbi decoding, span refinement, and a Faker-backed obfuscation engine. Same call on every host — Apple Silicon picks up MLX automatically; everywhere else uses this PyTorch checkpoint.

pip install -U "openmed[hf]"

from openmed import extract_pii, deidentify

text = (
    "Patient Sarah Johnson (DOB 03/15/1985), MRN 4872910, "
    "phone 415-555-0123, email [email protected]."
)

# Extract grouped entity spans
result = extract_pii(text, model_name="OpenMed/privacy-filter-multilingual")
for ent in result.entities:
    print(f"{ent.label:30s} {ent.text!r}  conf={ent.confidence:.2f}")

# De-identify with any of the supported methods
masked   = deidentify(text, method="mask",   model_name="OpenMed/privacy-filter-multilingual")
removed  = deidentify(text, method="remove", model_name="OpenMed/privacy-filter-multilingual")
hashed   = deidentify(text, method="hash",   model_name="OpenMed/privacy-filter-multilingual")

# Faker-backed locale-aware obfuscation, deterministic with consistent=True+seed
fake = deidentify(
    text,
    method="replace",
    model_name="OpenMed/privacy-filter-multilingual",
    consistent=True,
    seed=42,
)
print(fake.deidentified_text)

OpenMed/privacy-filter-multilingual-mlx* model names also work in the same extract_pii() / deidentify() calls — on a non-Apple-Silicon host they automatically fall back to this PyTorch checkpoint with a one-time warning. So you can ship MLX names in code and still run on Linux/Windows.

The OpenMed wrapper passes trust_remote_code=True for you, runs the model's own BIOES Viterbi decoder, and skips OpenMed's regex smart-merging (the model already produces clean spans).

Label space (54 categories)

The output space is O plus B- , I- , E- , S- for each of the 54 categories (4 × 54 + 1 = 217). The id2label mapping is shipped with the model.

Limitations & intended use

• Multilingual but uneven. Strongest on languages with rich PII training data (German, Spanish, French, Italian, Hindi, Telugu, English). CJK languages (Japanese, Korean, Chinese) and some morphologically-marked low-resource languages remain the main bottleneck on the current training mix.
• Synthetic training data. The AI4Privacy datasets are template-synthesized; real clinical notes, legal documents, and web text may show different surface forms. For high-stakes deployments, collect a domain-specific eval set and re-calibrate thresholds.
• Not a substitute for legal compliance review. Use alongside a governance layer (human review, deterministic regex pre-filters, etc.).
• Not a clinical PHI model. Healthcare-specific PHI and clinical entity training is planned as a separate branch.

Head initialization : opf 's default "copy-from-matching-base" head init. Of the 217 new BIOES classes, the few with exact base-vocabulary matches ( O , B/I/E/S-account_name , etc.) were copied directly; the rest were copied from semantically-adjacent coarse rows and fine-tuned end-to-end.

Router : base model has 128 MoE experts per layer with top-4 routing. Routers were kept trainable during full fine-tuning; no collapse was observed.

Credits & Acknowledgements

This model wouldn't exist without two open-source releases — sincere thanks to both teams:

• OpenAI for open-sourcing the Privacy Filter (architecture, modeling code, and opf training/eval CLI). Everything in this repo is a fine-tune on top of that release.
• AI4Privacy for releasing the multilingual PII masking datasets used as training data: pii-masking-200k , pii-masking-400k , open-pii-masking-500k-ai4privacy .

Additional thanks to the HuggingFace team for the transformers / huggingface_hub ecosystem this model ships through.

License

Apache 2.0.

Citation

If you use this model, please cite this model , the organization behind it ( OpenMed ), and the upstream base model + datasets:

@misc{openmed_privacy_filter_multilingual_2026,
  author       = {OpenMed},
  title        = {{OpenMed/privacy-filter-multilingual}: multilingual fine-grained PII extraction across 16 languages and 54 categories},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/OpenMed/privacy-filter-multilingual}}
}

@misc{openmed_2026,
  author       = {OpenMed},
  title        = {{OpenMed}: open models and resources for healthcare NLP},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/OpenMed}}
}

@misc{openai_privacy_filter_2025,
  author       = {OpenAI},
  title        = {{openai/privacy-filter}},
  year         = {2025},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/openai/privacy-filter}}
}

@misc{ai4privacy_pii_masking,
  author       = {AI4Privacy},
  title        = {{AI4Privacy PII Masking Datasets}},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/ai4privacy}}
}