{
  "model": "artifacts/tookabert-pii-4bit/mlx-custom",
  "dataset": "data/final_splits_audited/combined_clean",
  "split": "test",
  "rows": 2000,
  "max_length": 256,
  "batch_size": 8,
  "precision": 0.9823093626261336,
  "recall": 0.9845100172822121,
  "f1": 0.9834084588088615,
  "accuracy": 0.9962101153578585
}

Fixture/runtime verification:

json

{
  "status": "converted_mlx_int4",
  "weights": "artifacts/tookabert-pii-4bit/mlx-custom/weights.safetensors",
  "bits": 4,
  "group_size": 64,
  "max_length": 256,
  "verification": {
    "name": "mlx_int4",
    "shape": [
      2,
      256,
      39
    ],
    "argmax_match_rate_vs_unquantized_mlx": 0.921875,
    "max_abs_diff_vs_unquantized_mlx": 8.623085975646973,
    "mean_abs_diff_vs_unquantized_mlx": 0.14020052552223206
  }
}

Runtime Contract

Use this model behind the same production wrapper as the ONNX/CoreML releases:

sliding-window inference, usually max_length=256 and stride around 96;
offset-based span reconstruction;
whitespace trimming and overlap de-duplication;
deterministic regex/rule assists for email, phone, national ID, postal code, date, card number, and IMEI exclusion;
cue-word correction around Persian labels such as کد ملی, شماره تماس, کدپستی, and ایمیل.

"""Minimal MLX wrapper contract.

This repo includes a custom BERT token-classification MLX runtime script in the source project. Load weights.safetensors into the same module shape, tokenize with the bundled tokenizer, run sliding windows, then reconstruct spans from offsets and apply the same regex/rule postprocessing used by the ONNX/CoreML packages. """

Strongest Persian-script F1; trim whitespace spans in postprocessing.

Reza2kn/openmed-persian-pii-tookabert-large-mlx-int4

OpenMed Persian PII TookaBERT Large MLX INT4

Metrics

Runtime Contract

Reza2kn/
openmed-persian-pii-tookabert-large-mlx-int4