Bayanat-Labs/arabic-pii-eval

Arabic PII evaluation set (synthetic)

By Bayanat Labs. Sentences in Arabic and Arabic-English text with gold spans for personal data: Saudi, Emirati, Egyptian, Kuwaiti and Bahraini ID numbers, Arab IBANs and phone numbers, Saudi and UAE vehicle plates, e-mail addresses, URLs and person names. It was written to score arabic-pii-pack, a set of Arabic and Gulf recognisers for Microsoft Presidio, and it works for any detector that returns character spans. Try the recognisers in the browser: Bayanat-Labs/arabic-pii-redactor.

Every value is generated. Generated examples; not real personal data; any match to a real person is coincidental. ID numbers, IBANs and phone numbers pass the same structure and check-digit tests as real ones but were drawn at random; e-mail addresses and URLs use the reserved example.com, example.org and example.net domains; names combine common given and family names at random. Every row has synthetic: true.

This is an evaluation resource, not training data.

Config	Rows	What a row is
test	1,000	a sentence with its gold spans (1,383 spans; 200 sentences without personal data)
dev	300	the same, for tuning (413 spans)
results_per_entity	54	precision, recall and F1 per entity type for arabic-pii-pack 0.1.0 (rules only, and rules with the name model; exact and overlap matching)
wiki_false_positives	5,000	one Arabic Wikipedia paragraph id with the number and types of spans flagged in it (no text)

from datasets import load_dataset
test = load_dataset("Bayanat-Labs/arabic-pii-eval", "test", split="train")

The same sentences are in jsonl/ as JSON Lines, and the full metrics, per-sentence predictions and chart are in results/.

Fields

• id: test-0001 ... or dev-0001 ...
• text: the sentence, exactly as scored (it may contain right-to-left marks U+200F and left-to-right marks U+200E).
• spans: gold spans, each with start and end (Unicode code-point offsets into text, end exclusive), entity_type, text (the covered string), and where relevant country (ISO 3166-1 alpha-2 of an IBAN or phone number), phone_form (international, e164, 00, national, compact_national) and variant (rlm_inside, trailing_plus, parentheses, lower_case).
• has_pii: false for the sentences without personal data (look-alike numbers only: order and invoice numbers, amounts, dates, years, flight and tracking numbers).
• variety: msa (Modern Standard Arabic), gulf, egyptian or mixed_en (Arabic-English).
• digit_script: digits used in the sentence: western, arabic_indic, persian or none.
• template_id: the sentence template; synthetic: always true.

Entity types in the test split: PERSON 378, PHONE_NUMBER 330, IBAN_CODE 149, SA_NATIONAL_ID 95, EMAIL_ADDRESS 88, SA_IQAMA 81, EG_NATIONAL_ID 72, URL 38, KW_CIVIL_ID 36, AE_EMIRATES_ID 35, BH_CPR 31, SA_VEHICLE_PLATE 28, AE_VEHICLE_PLATE 22.

How it was made

eval/make_eval_set.py in the code repository fills 133 sentence templates (Modern Standard Arabic, Gulf and Egyptian phrasing, Arabic-English mixed) with values from bl_arabic_pii.fakes. Varieties are drawn 45/25/15/15 per cent; digits per sentence are Western (55%), Arabic-Indic (38%) or Persian (7%) in Arabic sentences; 20% of sentences have no personal data. Some values carry the surface variants listed above, and some numbers are preceded by a direction mark. Seeds: test 20261015, dev 7. The script rebuilds both files byte for byte.

The Arabic templates were written by Bayanat Labs; the Arabic text has not yet been reviewed by a native Arabic writer.

Results for arabic-pii-pack 0.1.0

Test split, all recognisers including the optional name model (CAMeLBERT MSA NER), threshold 0.5, exact span match; intervals from 1,000 bootstrap resamples of sentences:

Entity type	Gold spans	Precision	Recall	F1	F1 95% interval
SA_NATIONAL_ID	95	1.000	1.000	1.000	1.000-1.000
SA_IQAMA	81	1.000	1.000	1.000	1.000-1.000
AE_EMIRATES_ID	35	1.000	1.000	1.000	1.000-1.000
EG_NATIONAL_ID	72	1.000	1.000	1.000	1.000-1.000
KW_CIVIL_ID	36	1.000	1.000	1.000	1.000-1.000
BH_CPR	31	1.000	0.968	0.984	0.943-1.000
IBAN_CODE	149	1.000	1.000	1.000	1.000-1.000
PHONE_NUMBER	330	0.997	0.994	0.995	0.989-1.000
SA_VEHICLE_PLATE	28	1.000	1.000	1.000	1.000-1.000
AE_VEHICLE_PLATE	22	1.000	1.000	1.000	1.000-1.000
EMAIL_ADDRESS	88	1.000	1.000	1.000	1.000-1.000
URL	38	1.000	1.000	1.000	1.000-1.000
PERSON	378	0.937	0.820	0.874	0.844-0.901

With overlap matching, PERSON reaches precision 0.994, recall 0.870 and F1 0.928. None of the 200 sentences without personal data received a flag.

The format-checked recognisers score F1 0.984 to 1.000. The same team wrote the generators, the templates and the recognisers, so these numbers show format coverage (all three digit scripts, separators, direction marks, clitics), not accuracy on real documents.

False positives on Arabic Wikipedia. 5,000 paragraphs of 200-2,000 characters (1,729,163 characters) from 25 random row groups of wikimedia/wikipedia (20231101.ar, revision b04c8d1ceb2f): 0 identifier, IBAN, phone or plate flags at the default threshold 0.5, and 11 at 0.3 (10 phone-like, 1 plate-like). The wiki_false_positives config lists paragraph ids (<page id>:<paragraph index>) and flag types only; it contains no Wikipedia text. Wikipedia text is CC BY-SA 4.0 by its contributors.

Limitations

• Synthetic sentences are cleaner and more regular than real documents; scores here are an upper bound for format coverage.
• No free-text addresses, dates of birth, Qatari, Omani or Jordanian ID numbers, spoken-form numbers or sensitive categories.
• Names are common given and family names; mixed_en sentences also use names in Latin letters.

Licence

CC BY 4.0 (LICENSE). Attribute as: "Arabic PII evaluation set, Bayanat Labs (https://bayanatlabs.com), CC BY 4.0".

Citation

@misc{bayanatlabs_arabic_pii_eval_2026,
  author = {{Bayanat Labs}},
  title  = {Arabic PII evaluation set (synthetic)},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Bayanat-Labs/arabic-pii-eval}
}

العربية

لم يراجع النصَّ العربي بعدُ كاتبٌ عربي أصلي؛ والنص الإنجليزي هو المرجع.

مجموعة تقييم اصطناعية من بيانات لابز لاكتشاف البيانات الشخصية في النصوص العربية والمختلطة بالإنجليزية: 1,000 جملة اختبار و300 جملة تطوير مع مواضع البيانات الشخصية. القيم كلها مولَّدة وليست بيانات شخصية حقيقية، وأي تطابق مع شخص حقيقي محض مصادفة.

Contact

hello@bayanatlabs.com