The model was exported from PyTorch to ONNX format and dynamically quantized to INT8 using AVX2 instructions — matched to the AMD Ryzen 5 3500U instruction set. It achieves identical F1 to the PyTorch model with significantly lower CPU latency.

Performance

Metric	PyTorch (T4 GPU)	ONNX INT8 (Ryzen 5 CPU)
Overall F1	0.9916	0.9916
Latency p50	7.35ms	57.29ms
Latency p95	7.97ms	96.38ms
Within 150ms target	✓	✓

Zero accuracy loss from quantization — F1 preserved at 0.9916.

Hardware: CPU benchmark on ThinkPad AMD Ryzen 5 3500U (4 cores, no discrete GPU).

Detected Entity Types

PERSON, EMAIL, PHONE_NUMBER, ADDRESS, ACCOUNT_NUMBER, BANK_ACCOUNT_NUMBER, IBAN

Usage

python

from optimum.onnxruntime import ORTModelForTokenClassification
from transformers import AutoTokenizer

model = ORTModelForTokenClassification.from_pretrained("C-Ilyas/arabic-pii-detector-onnx")
tokenizer = AutoTokenizer.from_pretrained("C-Ilyas/arabic-pii-detector-onnx")

For the full inference pipeline (BIO decoding, redaction, confidence scores) see scripts/infer_onnx.py in the project repository.

How it was created

bash

python scripts/optimize_onnx.py \
    --model models/arabic-pii-detector \
    --output models/arabic-pii-detector-onnx

Uses AutoQuantizationConfig.avx2(is_static=False, per_channel=False) — dynamic INT8 quantization with AVX2 kernels.

PyTorch Model

The original PyTorch model with full documentation, training details, and per-class metrics is at: C-Ilyas/arabic-pii-detector