mabahboh/sitr-arabic-pii

سِتر · Sitr

An Arabic-first PII detection dataset for Saudi Arabia, the Gulf and the Levant

Character-level spans for 43 personal-data types across 17,899 Arabic examples — regional dialects, code-switching, call-centre transcripts, and the long documents people paste into AI assistants.

---

Why this exists

General-purpose PII filters underperform badly in Arabic. Independent evaluation of the widely used openai/privacy-filter reports roughly 0.198 F1 on Arabic — about 0.40 on person names and 0.49 on addresses — against far stronger numbers in English. Multilingual alternatives cover Arabic nominally, but their strength sits in European and Indic languages.

The gap is not only Arabic. It is Saudi and Gulf specific, and no volume of general Arabic text closes it:

Identifier	Why generic models miss it
National ID / Iqama	10 digits, leading 1 or 2, Luhn-checked — indistinguishable from any other long number without the rule
Saudi & GCC IBANs	SA + 22 digits; bank code carries meaning
Mobile numbers	05…, +966…, 00966… and bare 5… all in circulation
National Address	Short code, building number, postal code and secondary number as one unit
Commercial registration, title deeds, Ejar contracts	Domain-specific formats absent from Western corpora
Hijri dates	1445/03/22 sits in the same field as 2024-10-06
Gulf & Syrian documents	Emirates ID, Kuwaiti Civil ID, Qatari QID, Bahraini CPR, Omani Civil Number, Syrian National ID
Arabic-Indic digits	٠٥٥١٢٣٤٥٦٧ — invisible to [0-9]
Numbers spoken as words	صفر خمسة خمسة… in call transcripts — no regex and no current model catches these

This dataset targets exactly those.

---

At a glance

Rows	17,899
Labelled spans	40,803
Entity types	43 fine-grained, mapped onto 8 base classes
Hard negatives	3,620 rows containing no PII at all
Multi-turn dialogue rows	7,428
Mean spans per row	2.28
Languages	Arabic (MSA + dialects), English, Arabizi, code-switched
Splits	train / validation / test — leakage-free by construction

---

Quick start

from datasets import load_dataset

ds = load_dataset("mabahboh/sitr-arabic-pii")

row = ds["train"][0]
print(row["text"])         # هذا الهوية الوطنية 1429096553
print(row["entities"])     # [{'start': 19, 'end': 29, 'label': 'national_id', ...}]
print(row["secret_text"])  # هذا الهوية الوطنية {national_id}

Spans are plain character offsets, so aligning them to any tokeniser is direct:

enc = tokenizer(row["text"], return_offsets_mapping=True)
for (tok_start, tok_end), token in zip(enc["offset_mapping"], enc.tokens()):
    label = next((e["label"] for e in row["entities"]
                  if tok_start >= e["start"] and tok_end <= e["end"]), "O")

---

Data fields

Field	Type	Description
text	string	The text exactly as a filter would receive it.
entities	list<{start:int, end:int, label:string, text:string}>	Character-level spans. start inclusive, end exclusive, as Python string indices into text.
type	string	The row's primary category — the first entity's label, or negative.
secret_text	string	text with every span replaced by {label} — the masked output a filter should produce.

Examples

A single identifier in a short message:

{
  "text": "هذا الهوية الوطنية 1429096553",
  "entities": [
    {"start": 19, "end": 29, "label": "national_id", "text": "1429096553"}
  ],
  "type": "national_id",
  "secret_text": "هذا الهوية الوطنية {national_id}"
}

A dense record of the kind pasted into a chat assistant, where the request line is clean and every identifier sits inside the pasted block:

{
  "text": "راجع لي هذا الكشف:\n\nالرقم الوظيفي: E-52316 | الجوال: 0559143827\nالرقم الوظيفي: SA-1034 | الجوال: 0532477660",
  "type": "employee_id",
  "secret_text": "راجع لي هذا الكشف:\n\nالرقم الوظيفي: {employee_id} | الجوال: {private_phone}\nالرقم الوظيفي: {employee_id} | الجوال: {private_phone}"
}

---

Label taxonomy

Labels are fine-grained, and label_map.json folds each onto one of the eight privacy-filter base classes. Train at whichever level suits you: the fine labels carry the detail, the base classes carry the compatibility.

Base class	Fine-grained labels
account_number	national_id, iqama, account_number (IBAN), gulf_iban, credit_card, passport_number, boundary_number, visitor_number, commercial_registration, deed_number, contract_number, medical_record_number, insurance_policy, vehicle_plate, employee_id, emirates_id, kuwait_civil_id, qatar_qid, bahrain_cpr, oman_civil_id, syria_national_id
private_phone	private_phone, gulf_phone, syria_phone, spoken_phone
private_address	private_address, national_address, short_address, building_number, postal_code
secret	otp_code, password, card_pin, card_cvv, api_key, access_token, db_connection, spoken_otp
private_email	private_email
private_date	private_date
(unmapped)	medicine_name — see Notes for users

privacy-filter has no government-ID class, so identifiers fold onto account_number, the closest of the eight.

---

Coverage

Four registers, each chosen because a deployed filter meets it:

1. Contact-centre messages — complaints, follow-ups and transcribed calls in Najdi, Hijazi, Eastern, Southern, Gulf, Egyptian and Levantine dialects.
2. Multi-turn agent ↔ customer dialogues, marked الموظف: / العميل:.
3. Text pasted into AI assistants — a short request wrapping a pasted document. The identifiers sit inside the pasted block while the request line stays clean, which is precisely the discrimination a client-side filter has to make.
4. Workplace records — payroll tables, customer listings, and source code carrying API keys and database connection strings.

Surface variation is present throughout rather than normalised away: missing hamza, ة/ه and ى/ي alternation, absent punctuation, elongation (وااايد), typographic errors, Arabic-Indic digits, and identifiers written flush against the preceding word (أيجار2246817551) — the case where a word-boundary match cannot fire in Arabic.

---

Data quality

Three properties hold across every row, and each is verified rather than assumed:

Span offsets are exact. Every entities[i] satisfies text[start:end] == text_field for all 40,803 spans, verified on the released files.

Identifiers are format-valid. Saudi National ID and Iqama numbers carry a correct Luhn check digit; every IBAN passes mod-97; payment card numbers pass Luhn; Emirates IDs pass Luhn.

Checksum validity is deliberately mixed. Numbers captured from speech and manual entry are frequently mistyped in practice, so a proportion of identifiers fail their checksum by design. A corpus in which every number validates cleanly would flatter a checksum-based rule layer far beyond its behaviour in production.

Splits are leakage-free

Splits are partitioned by document family, not sampled row-wise. Rows sharing a common structure are placed together, so no near-duplicate of a test row appears in training. Sampling at random instead would place textual twins on both sides and inflate every reported score — on a corpus of this shape, by roughly 3.8% of training rows. Here that figure is zero, verified on the released files using the published text alone.

Split	Rows
train	14,288
validation	1,818
test	1,793

---

Intended use

• Training and evaluating Arabic and Gulf PII detection and de-identification models
• Building client-side privacy filters for LLM chat interfaces
• Benchmarking redaction gateways in a PDPL context

Out of scope

• Not a compliance certification. A model trained on this reduces exposure; it does not establish that any system is anonymous or PDPL-compliant.
• Not a source of personal data. Every identifier is fictitious and corresponds to no real person, account or system.

---

Notes for users

Stated plainly, because they affect how results should be read.

1. Class balance. account_number absorbs 21 fine-grained types and dominates the base-class distribution. Rebalance before training on the collapsed classes.
2. medicine_name is intentionally unmapped. A drug name is sensitive health information but it is not an identifier — it does not point at an individual. Folding it into the eight would change the task from de-identification to sensitive-content detection. Treat it as a deliberate choice.
3. private_person is not covered. Person names are absent as a labelled class, which is a meaningful gap: names are where general-purpose models fail hardest in Arabic.
4. Checksum coverage varies by document. Saudi ID and Iqama (Luhn), all IBANs (mod-97) and payment cards (Luhn) are verifiable. Kuwaiti, Qatari, Bahraini, Omani and Syrian identifiers carry correct digit counts and leading blocks but no verifiable check digit. Syria is absent from the ISO IBAN registry, so no Syrian IBAN appears.
5. A manually reviewed benchmark is a separate artifact. Scores on this corpus measure fit to this distribution. An independently collected, hand-reviewed evaluation set is the stricter measuring stick and is not part of this release.

---

Licence

Released under the Apache License 2.0. You may use, modify and redistribute it, including commercially, with attribution and the notice preserved.

No personal data is included. Records derived from internal customer material are excluded from this release and retained for internal evaluation only.

---

Citation

@misc{sitr_arabic_pii,
  title  = {Sitr: An Arabic-First PII Detection Dataset for Saudi Arabia,
            the Gulf and the Levant},
  author = {Mohammad Bhbouh},
  year   = {2026},
  url    = {https://huggingface.co/datasets/mabahboh/sitr-arabic-pii}
}

---

_{Built for a privacy layer that runs on the user's own device — so the model never sees the name in the first place.}