PII-Redact-General-oQ4

This model was quantized using oQ (oMLX v0.7.0) mixed-precision quantization.

Quantization details

• Model type: llama
• Bits: 4
• Group size: 64
• Format: MLX safetensors

Quantization of OpenPipe/PII-Redact-General (oQ4).

Usage

This quant is meant to be served by oMLX and driven through the PII-redaction fork, which pairs it with the Name model and handles tagging, redaction, and fake-data replacement:

https://github.com/gwyntel-git/pii-redact-mlx

pip install "git+https://github.com/gwyntel-git/pii-redact-mlx"

# list the models an oMLX server is serving
pii-redact list-models --omlx-url http://localhost:27473/v1

# redact a trace/JSONL file
pii-redact convert-traces input.jsonl output.jsonl \
  --backend omlx --omlx-url http://localhost:27473/v1 --auto-max-tokens

The fork also provides a transformers/torch backend, concurrent inference, and a labelled validation harness. Validated on an 18-document labelled set: recall matches the full-precision model, F1 0.816 vs 0.833 (precision 0.870 vs 0.909).