hehehe10101010/anonymous-pii-model

Anonymous model release

BIOES token-classification model for multilingual PII span extraction (nine labels, 37 tags), released anonymously for double-blind review. It accompanies the anonymous submission.

• adapter/ + classifier.pt: LoRA adapter (rank 128) and tagging head applied to LiquidAI/LFM2.5-Encoder-350M-PII-Detector at load time (the evaluated path).
• model.safetensors: self-contained merged weights (a convenience form; its span differences against the adapter form are recorded in config.json under release.parity).
• pii_decode/: Viterbi BIOES decoder with character offsets.
• modeling_pii.py: loader and extractor.

from modeling_pii import PiiExtractor
model = PiiExtractor.from_pretrained(".", weights="adapter")  # or weights="merged"
spans = model.extract("Contact Jane Doe at jane@example.com")

Requires torch, transformers, peft, safetensors. Labels: address, company_name, date, email_address, human_name, id_number, phone_number, private_url, secret. Licenses: see LICENSE, LICENSE-LFM-OPEN-V1.0, NOTICE.