OpenAI PII Filter Retrained

OpenAI privacy-filter has been fully fine-tuned on NVIDIA Nemotron-PII, expanding the existing 8 broad PII categories into 55 fine-grained labels such as first_name, last_name, medical_record_number, credit_debit_card, and ssn, released as OpenMed/privacy-filter-nemotron. The base model is a 1.4B parameter MoE (50M active per token) and retains the BIOES token classification head.

• Training data: the full 100K rows of the nvidia/Nemotron-PII train split
• Validation data: 10K rows sampled from the test split (label-stratified, at least 229 entities per label)
• Training setup: opf train, full fine-tune, AdamW, lr 1e-4, 5 epochs, bf16, weight decay 0.0
• Output space: 221 BIOES classes (1 O class plus B/I/E/S for each of the 55 categories)

In evaluation, based on opf eval --decode-mode viterbi --eval-mode typed --span-metrics-space char, it achieved macro B-F1 0.9533 and token accuracy 0.9910. 46/55 labels scored F1 0.90 or higher, and there were no weak labels.

mac_address, biometric_identifier, bank_routing_number, and credit_debit_card showed F1 scores in the 0.99 range, while occupation 0.727, language 0.804, state 0.829, and gender 0.841 were relatively lower. It can be used directly via OpenMed's extract_pii() / deidentify() and the opf redact CLI, and MLX BF16 and 8-bit variants for Apple Silicon are also provided.