This repo contains the verified ONNX INT4 artifact only. Other requested runtime formats were tested or inspected and are not uploaded here unless they pass runtime verification.

Format	Status	Note
ONNX INT4	verified	`model.onnx` loads in ONNX Runtime CPU and was evaluated on held-out rows.
CoreML 4-bit	blocked	CoreMLTools produced a package, but FP32 CoreML prediction did not match PyTorch closely enough, so it was not published.
AWQ	blocked	Standard AWQ tooling targets causal LLMs, not BERT token-classification encoders.
MLX DWQ	blocked	Current `mlx-lm` DWQ path is LLM-oriented and did not expose a verified BERT token-classifier conversion path.
LiteRT 4-bit	blocked	LiteRT/TFLite has no verified 4-bit BERT token-classification path in this workflow.

Metrics

Original dense held-out test F1: 0.9800

ONNX INT4 held-out slice eval (test, first 2,000 rows, max_length=256):

json

{
  "model": "artifacts/tookabert-pii-4bit/onnx/model.int4.matmulnbits.onnx",
  "dataset": "data/final_splits_audited/combined_clean",
  "split": "test",
  "rows": 2000,
  "max_length": 256,
  "batch_size": 16,
  "precision": 0.9828071072478589,
  "recall": 0.9842539845100173,
  "f1": 0.9835300137516391,
  "accuracy": 0.9961062829019095
}

Ad hoc edgecase probes are included under reports/.

Recommended Inference Wrapper

Do not run this model as a single truncated pass over long documents. Production use should wrap the ONNX model with:

Sliding windows: tokenize with max_length=256 or 512, stride=96 or similar, and merge chunk outputs by original character offsets.
Span cleanup: trim leading/trailing whitespace, merge adjacent spans with the same label, and de-duplicate overlapping chunk predictions.
Regex/rule assists: add high-precision detectors for emails, Iranian mobile/phone numbers, national IDs, postal codes, dates, Shetab/card numbers, and IMEI exclusion.
Context correction: use nearby Persian cue words such as کد ملی, گواهینامه, گذرنامه, کدپستی, شماره تماس, and ایمیل to correct ambiguous numeric labels.
Obfuscation handling: normalize forms like name [dot] x [at] y, spaced phone digits, and verbal Persian phone numbers before or after model inference.

Minimal Python ONNX Runtime sketch:

python

import onnxruntime as ort
from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained(".")
session = ort.InferenceSession("model.onnx", providers=["CPUExecutionProvider"])

def run_window(text, max_length=256):
    enc = tokenizer(
        text,
        return_offsets_mapping=True,
        return_tensors="np",
        truncation=True,
        padding="max_length",
        max_length=max_length,
    )
    offsets = enc.pop("offset_mapping")[0]
    if "token_type_ids" not in enc:
        import numpy as np
        enc["token_type_ids"] = np.zeros_like(enc["input_ids"])
    logits = session.run(None, {k: v.astype("int64") for k, v in enc.items()})[0][0]
    # Convert argmax labels to spans, trim whitespace, merge overlaps, then apply regex/rule postprocessing.
    return logits, offsets

Known Edge Cases

Long documents require sliding-window inference.
Obfuscated contacts and verbal/spaced phone numbers need deterministic normalization/rules.
IMEI-like device IDs can look like card numbers; validate card numbers before masking as CREDITCARDNUMBER.
Compact addresses may confuse CITY and STREET; use address cue-word postprocessing.

Best formal F1 and strong Persian-script robustness; spans often include leading whitespace, so trim postprocessing is mandatory.

Files

model.onnx: 4-bit ONNX Runtime MatMulNBits model.
tokenizer/config files from the trained checkpoint.
verification.json: fixture-level ONNX verification.
reports/: ad hoc edgecase reports.

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

OpenMed Persian PII TookaBERT Large ONNX INT4

Status

Metrics

Recommended Inference Wrapper

Known Edge Cases

Files

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