Fine-tuned token classification model that detects Personally Identifiable Information (PII) in Arabic, English, and mixed Arabic/English text. Covers Modern Standard Arabic, Gulf, Egyptian, and Algerian dialects.

Detected Entity Types

Tag	Description
`PERSON`	Full name / person name
`EMAIL`	Email address
`PHONE_NUMBER`	Phone number
`ADDRESS`	Physical address
`ACCOUNT_NUMBER`	Account number (internal / short)
`BANK_ACCOUNT_NUMBER`	Bank account number (longer numeric)
`IBAN`	International Bank Account Number

Base Model

aubmindlab/bert-base-arabertv02 — selected after evaluating multiple alternatives:

~135M parameters (12 transformer layers, hidden size 768)
Pretrained on large Arabic corpora covering MSA and dialectal Arabic
Handles code-switching with English natively
CPU ONNX INT8 expected p95 ~60–90ms , well within the <150ms latency budget
Has safetensors format ✓

Why not bert-large? aubmindlab/bert-large-arabertv02 (340M params) was tested first , but failed the CPU latency target (p95=239ms with ONNX INT8 on AMD Ryzen 5 3500U).

Why not mdeberta-v3-base? DeBERTa's disentangled attention adds CPU overhead , estimated p95 150–250ms, borderline at best. Not worth the complexity.

Training Data

Synthetic training data generated by scripts/prepare_data.py:

~130 sentence templates covering single entity, multi entity, code-switched, and Algerian dialect cases
Diverse entity value pools (scripts/entity_pools.py):
- Names: Arabic (MSA + Gulf + Egyptian + Algerian), Berber-influenced.
- Phones: Egypt, Saudi, UAE, Algeria (+213 / 05xx/06xx/07xx)
- IBANs: valid mod-97 checksums for EG/SA/AE/JO/KW/QA/BH/DZ
- Emails: includes .dz, .fr, yahoo.fr, hotmail.fr, univ-alger.dz
- Addresses: 20 Algerian cities, Algerian districts and street names, plus Egypt/Saudi/UAE
25% negative examples : hard distractors: tax IDs (NIF), SWIFT codes, IMEI numbers, decree numbers, invoice refs, customs codes

Evaluation split strategy: All three splits (train/val/test) use structurally distinct held-out templates. Val and test templates are of equal difficulty ; a mix of simple, ambiguous, and hard multi-entity cases (up to 4 entities), including the ACCOUNT_NUMBER vs BANK_ACCOUNT_NUMBER confusion pair in both pools. This ensures val F1 during training is honest and comparable to test F1.

Split	Examples	Templates
train	50,000	~110 (from TEMPLATES pool)
validation	2,000	21 (VAL_ONLY_TEMPLATES, held-out)
test	2,000	21 (TEST_ONLY_TEMPLATES, held-out)

Label Schema

BIO tagging scheme with 15 labels:

O, B-PERSON, I-PERSON, B-EMAIL, I-EMAIL,
B-PHONE_NUMBER, I-PHONE_NUMBER, B-ADDRESS, I-ADDRESS,
B-ACCOUNT_NUMBER, I-ACCOUNT_NUMBER,
B-BANK_ACCOUNT_NUMBER, I-BANK_ACCOUNT_NUMBER,
B-IBAN, I-IBAN

Training Configuration

Training hardware: NVIDIA Tesla T4 GPU (Google Colab)

Hyperparameter	Value	Reasoning
Base model	aubmindlab/bert-base-arabertv02
Epochs	12 (early stopping)	Ceiling ,early stopping finds the peak
Early stopping patience	3	Cosine schedule can dip before recovering
Batch size	32	bert-base fits larger batches; more stable gradients
Learning rate	3e-5	bert-base needs higher lr than bert-large
LR scheduler	Cosine	Stays near peak lr longer, then decays smoothly — outperforms linear on NER by 0.5–1.5 F1 points
Warmup ratio	0.06	Short warmup; over-warming wastes steps with 50K examples
Weight decay	0.01	Standard for bert-base; 0.1 (used for bert-large) caused underfit
Label smoothing	0.0	bert-base is underconfident; smoothing hurt recall
Max sequence length	256
Optimizer	AdamW
fp16	true (GPU training)
Seed	42

Key insight: hyperparameters differ deliberately from bert-large. bert-large (340M params) needed heavy regularization (weight_decay=0.1, dropout=0.2, label_smoothing=0.1) to prevent overfitting synthetic data. Applying those same settings to bert-base (135M params) caused underfit , the smaller model needs lighter regularization and a higher learning rate to reach its capacity ceiling.

Preprocessing

Arabic-Indic digits (٠١٢٣...) normalized to Western digits (0123...) before tokenization 1:1 character mapping preserves original offsets
Fast (Rust-based) tokenizer for speed

Postprocessing

BIO decoding to spans
IBAN mod-97 checksum and email format checks adjust confidence (never override model prediction)
Trailing whitespace/punctuation trimmed from entity spans

Performance

Training & benchmark hardware: NVIDIA Tesla T4 GPU (Google Colab)

Evaluated on 2000 examples using held-out templates never seen during training:

Class	Precision	Recall	F1	Support
EMAIL	1.000	1.000	1.000	283
PERSON	0.996	0.996	0.996	516
IBAN	0.991	1.000	0.995	317
PHONE_NUMBER	0.994	0.991	0.993	350
BANK_ACCOUNT_NUMBER	0.968	1.000	0.984	273
ACCOUNT_NUMBER	0.990	0.958	0.974	214
ADDRESS	0.990	0.990	0.990	296
Overall			0.9916

Metric	Before fine-tuning	PyTorch (T4 GPU)	ONNX INT8 (Ryzen 5 CPU)
Overall F1	0.0006	0.9916	0.9916
Val F1	—	0.9682	—
Precision	—	0.9907	0.9907
Recall	—	0.9924	0.9924
Latency p50	—	7.35ms	57.29ms
Latency p95	—	7.97ms	96.38ms
Within 150ms	—	✓	✓

Hardware notes:

Training + GPU benchmark: NVIDIA Tesla T4 (Google Colab)
CPU benchmark: ThinkPad AMD Ryzen 5 3500U, 4 cores, no discrete GPU — typical CPU-only deployment machine
ONNX quantization uses AVX2 config (matches Ryzen 5 3500U instruction set , not AVX-512 which would silently fall back to a slower path)

On the val/test F1 gap: Val F1 (0.9682) and test F1 (0.9916) both use structurally distinct held-out templates of equal difficulty. The 2.3-point gap reflects natural random variation in generated entity values across the two pools — not data leakage. In real-world deployment expect performance closer to val F1 (0.9682) since real text contains noise not present in synthetic data.

Intended Use

Redaction pipelines for Arabic customer support data, chat logs, documents
Pre-processing before sending text to LLMs
Privacy compliance workflows in North Africa (Algeria, Morocco, Tunisia) and the broader Arab world

Limitations

Synthetic data domain gap: real text has OCR errors, dialectal variations, and structural patterns not covered by templates
ACCOUNT_NUMBER vs BANK_ACCOUNT_NUMBER: distinguished by length and context — borderline cases may be confused
No PII types beyond the 7 listed: national IDs, passport numbers, credit cards are NOT detected
Maximum input length: 256 tokens

How to Use

python

from transformers import AutoTokenizer, AutoModelForTokenClassification

tokenizer = AutoTokenizer.from_pretrained("C-Ilyas/arabic-pii-detector")
model = AutoModelForTokenClassification.from_pretrained("C-Ilyas/arabic-pii-detector")
# See scripts/infer.py for full pipeline with BIO decoding and redaction

bash

# CLI
python scripts/infer.py --text "اسمي محمد أحمد ورقم تليفوني 01012345678" --pretty
python scripts/infer.py --text "راني سفيان بوزيد ورقمي 0698123456 من الجزائر العاصمة" --pretty

Evaluation

See results/after_finetuning_metrics.json for full per-class metrics, results/latency_report.json for latency profiling, and results_onnx/ for CPU ONNX benchmark results.