lan1983/pii-guard-tagger

PII Guard tagger (private, personal use)

Token classifier for Indian PII (16 types: NAME, ADDRESS, PHONE, EMAIL, UPI, AADHAAR, PAN, GSTIN, CARD, BANK_ACCT, PASSPORT, VOTER_ID, VEHICLE, DOB, IP, DL) in English, Hinglish and South Indian languages written in Latin script. Used with the rules + format checks in pii-guard.

Licence: fine-tuned from iiiorg/piiranha-v1-detect-personal-information (CC BY-NC-ND 4.0). Non-commercial, personal use only; this derivative must not be shared or deployed commercially. For commercial use, retrain from microsoft/deberta-v3-base (MIT) with the repo's pii.train --base.

• Training: 15,000 synthetic examples, 3 epochs, full fine-tune; best validation loss 0.0001
• Engine leak (PII characters left visible): 0.4–0.6% Ylemis India-PII, 2–5% Maskara, 0.3% own hand-written set
• Files: model.safetensors (PyTorch), onnx/model.onnx (serving, identical output, ~9 ms on CPU)