nym-pii-multilingual

Multilingual PII token-classification model for the nym anonymization CLI — a fine-tune of jhu-clsp/mmBERT-base (ModernBERT architecture, 1,800-language pretraining) on fully synthetic, exactly-labeled PII data.

40 entity types / 81 BIO labels across ~23 languages and 6 scripts (Latin, Cyrillic, CJK, Korean, Arabic, Devanagari): names, dates of birth, emails, phones, street addresses, cities, government/medical/financial IDs, credit cards, IBANs, API keys, passwords, and more. Trained with ~20% OCR-style corruption (o→0, rn→m, typos, spacing), so it catches PII that scanning mangles and exact regex misses.

Files

File	Use
model.safetensors	PyTorch / transformers (fine-tuning, GPU inference)
model.onnx	ONNX fp32 — nym's default (accuracy-first)
int8/	ONNX dynamic int8 variant: 309 MB vs 1.2 GB and ~3-4× faster on CPU, at a measured accuracy cost (see below). Use via token_model = "Wismut/nym-pii-multilingual/int8"

Benchmarks

Cross-dataset (out-of-distribution for this model), span-level, label-agnostic matching on 1,000 examples of ai4privacy/pii-masking-300k, via nym bench:

Model	Precision	Recall	F1
regex only	79.2	27.8	41.2
Rampart (MiniLM, in-distribution)	80.9	68.5	74.2
OpenMed-small (DeBERTa)	77.8	74.1	75.9
this model	80.6	79.0	79.8

Held-out synthetic test set (in-distribution): P 99.0 / R 99.3 / F1 99.1.

int8 trade-off (same ai4privacy protocol, n=200): fp32 F1 81.0 → int8 76.7 (recall 79.9 → 72.6; hits person/time/zip recall hardest). Pick fp32 for accuracy, int8 for constrained machines.

Use with nym

This is nym's default token-classification model — with NER enabled it is downloaded and cached automatically. Explicitly:

[ner]
enabled = true
backend = "tokens"          # or "both" to also run GLiNER
token_model = "Wismut/nym-pii-multilingual"
threshold = 0.5

Use with transformers

from transformers import pipeline
nlp = pipeline("token-classification", model="Wismut/nym-pii-multilingual",
               aggregation_strategy="simple")
nlp("Patient John Smith, SSN 123-45-6789, lives at Hauptstraße 42, München.")

Training

• Data: Wismut/nym-pii-multilingual-data — 69,000 synthetic examples (62,100 train), generated by an LLM writing [LABEL]-placeholder templates across a (language × topic × style × flavor) rubric, filled by Faker with locale-matched values so character-level labels are exact by construction; ~13% PII-free negatives; ~20% OCR/typo noise. Generation pipeline: scripts/datagen in the nym repo.
• 3 epochs, lr 3e-5, batch 32, max_length 256, bf16, single RTX A6000 (~17 min). Training script: scripts/train_ner.py.
• ONNX export: opset 18 via optimum; int8 via onnxruntime dynamic quantization (functional — unlike DeBERTa-v2 it does not collapse — but with the accuracy cost quantified above).

Limitations

• Synthetic training data: strong on structured/semi-structured PII patterns; free-prose edge cases may differ from human-annotated data.
• Labels are English-named regardless of text language (by design — they map to nym's pattern names).
• No claim of HIPAA/GDPR compliance by itself; it is a detection component.

License

MIT (inherited from mmBERT-base). Synthetic training data contains no real personal information.