Mozilla/smart-pii-redactor-v1-onnx

smart-pii-redactor v1 (ONNX)

Token classifier (XLM-R MiniLM, 6 layers, 20k vocab) that finds identifiers in chat messages. Labels: PERSON, STREET_ADDRESS, DOB, ID_IN_CONTEXT, SECRET_IN_CONTEXT, USERNAME (replace); ORG, PRODUCT, LOCATION, DATE, PUBLIC_NAME (keep). Max 256 tokens per window; the caller splits longer text (stride 32).

Recommended: q4 (dtype: "q4", onnx/model_q4.onnx, 14.8 MB). It loads in older ONNX Runtime builds too, and costs about 0.1 points of leak recall against fp32.

File	transformers.js dtype	Size	Test leak recall	Keep retention
onnx/model_q4.onnx	q4 (4-bit weights, recommended)	14.8 MB	0.9788	0.9497
onnx/model_quantized.onnx	q8 (8-bit weights)	20.1 MB	0.9800	0.9510
onnx/model.onnx	fp32	74.3 MB	0.9800	0.9522
onnx/model_uint8.onnx	uint8 (dynamic; fallback)	18.8 MB	0.9654	0.9530

q8 is stored as model_quantized.onnx, the name transformers.js expects for dtype: "q8". It needs a recent ONNX Runtime (8-bit MatMulNBits) and does not load in onnxruntime-web 1.22. uint8 is only a fallback and loses about 1.5 points of leak recall.

All quantized files keep the word embeddings as int8. Scores: v1 on the 10,569-message test set (the project's own test messages plus held-out public rows, 8 languages), model alone: spans of the fixed-format types the rules own (emails, phones, cards, IBANs, ...) are not scored. Inputs: input_ids, attention_mask. Output: logits (batch, seq, 23).

Fine-tuned from nreimers/mMiniLMv2-L6-H384-distilled-from-XLMR-Large (MIT, Microsoft unilm). Trained on smart-pii-redactor-data (385,669 messages); its card lists the sources and their licences.