Bayanat-Labs/gcc-format-validators

GCC Format Validators and Arabic ITN/TN Test Pairs · مراجع الصيغ الخليجية واختبارات التحويل بين المنطوق والمكتوب

2,000 synthetic spoken/written test pairs in 10 classes, 3 varieties (MSA, Saudi/Gulf, Egyptian) and 2 digit scripts, plus reference tables for identity numbers, IBANs, phone numbering plans, Saudi plates, the Umm al-Qura calendar (55,991 days, 1343-1500 AH) and Arabic number and currency words. Version v0.1.0 (2026-10-15) · Validators, scorer and build scripts: Bayanat-Labs/gcc-format-validators (Apache-2.0) Published by Bayanat Labs (bayanatlabs.com), which is not affiliated with Space42 or the ARBML bayanat library.

1. What is it

Speech recognisers and text-to-speech front ends have to move between how numbers are said and how they are written: «خمسة وعشرون ألفًا وخمسمائة ريال» and 25,500 ريال, «الساعة ثلاث ونص» and 3:30, an IBAN read digit by digit and SA44 2000 .... This dataset gives you:

• Test pairs for inverse text normalisation (ITN, spoken to written) and text normalisation (TN, written to spoken), each with the accepted written and spoken variants, so systems can be scored per class, variety and digit script.
• Reference tables for the formats those pairs contain, with the evidence for each rule, and the same tables bundled in a small Python package of validators.

All identifiers, IBANs and phone numbers are random values that pass the format checks. Generated examples; not real personal data; any match to a real person is coincidental.

2. What is in this release

config	rows (split)	licence
itn_tn_pairs	dev 600, test 1,400	CC BY 4.0
formats	table 7	CC BY 4.0 (format facts CC0 1.0)
format_examples	table 147	CC BY 4.0
iban_registry	table 21	CC BY 4.0; structure data from schwifty (MIT)
phone_plans	table 22	Apache-2.0 (libphonenumber metadata)
plates_sa	table 17	CC BY 4.0 (facts CC0 1.0)
eg_governorates	table 28	CC BY 4.0 (facts CC0 1.0)
hijri_gregorian	table 55,991	CC BY 4.0; dates from hijridate (MIT)
hijri_month_starts	table 1,896	CC BY 4.0; dates from hijridate (MIT)
hijri_icu_disagreements	table 22,058	CC BY 4.0
number_words	table 8,404	CC BY 4.0; words_cldr Unicode License V3
currency_units	table 3	CC BY 4.0; cldr_* Unicode License V3

Reference tables use a single split named table. CSV copies of every table are in tables/, and tables/formats.json holds the format table as one JSON file.

3. Quick start

from datasets import load_dataset

pairs = load_dataset("Bayanat-Labs/gcc-format-validators", "itn_tn_pairs", split="test")
days = load_dataset("Bayanat-Labs/gcc-format-validators", "hijri_gregorian", split="table")

Validators, amounts in words and the scorer (Python 3.9+, no dependencies):

pip install "git+https://github.com/Bayanat-Labs/gcc-format-validators@v0.1.0"
gcc-formats validate sa-id 1000000008
gcc-formats validate phone-AE "+971 50 123 4567"
gcc-formats tafqeet 25500.50 --currency SAR      # خمسة وعشرون ألفًا وخمسمائة ريال وخمسون هللةً
gcc-formats hijri 1447-09-01                     # 2026-02-18
gcc-formats score --gold test.jsonl --pred predictions.jsonl --direction itn

A registry install (pip install gcc-format-validators) will work once the package is published.

4. ITN/TN test pairs

Each row has pair_id, class, subclass, variety (msa, gulf, egy), digit_script of the written form (western, arabic_indic), country (for format-specific rows), spoken, spoken_variants, written, written_variants, template_id, generator, seed, synthetic (always true), review_status and split.

class	MSA	Saudi/Gulf	Egyptian	total	Western digits	Arabic-Indic digits
cardinal	67	67	66	200	101	99
ordinal	67	67	66	200	101	99
money	67	67	66	200	101	99
date	67	67	66	200	101	99
time	67	67	66	200	101	99
telephone	67	67	66	200	101	99
id	67	67	66	200	101	99
iban	67	67	66	200	200	0
percent	67	67	66	200	101	99
measure	67	67	66	200	101	99
all	670	670	660	2000	1109	891

Splits: dev 600, test 1400 (stratified by class, variety and digit script; seeded). 82 templates. IBANs are written with Western digits only.

Subclasses: money (AED 46, EGP 86, SAR 68); date (gregorian 100, hijri 100); telephone (ae_mobile 48, eg_mobile 89, sa_mobile 63); id (ae_emirates_id 60, eg_national_id 94, sa_national_id 21, sa_resident_id 25); iban (ae_iban 46, eg_iban 96, sa_iban 58); measure (c 34, cm 34, kg 32, km 32, l 31, m 37).

class	variety	spoken	written
money	gulf	ستمية وثمانية درهم	608 درهم
date	msa	الخامس من ذي القعدة	5 ذو القعدة
time	egy	الساعة تلاتة إلا عشرة الضهر	2:50 م
telephone	gulf	زائد تسعة سبعة واحد خمسة خمسة خمسة خمسة ثلاثة ثمانية تسعة ستة اثنين	+971555538962
percent	egy	سبعة وتمانين في المية	87%

Review status. Every row has review_status = "not_reviewed_by_native_speaker" (2,000 of 2,000 rows). The MSA forms follow written grammar rules (gender polarity for 3-10, dual, accusative singular after 11-99, genitive singular after hundreds and thousands); the Saudi/Gulf and Egyptian forms use one common spelling for each word, and other spellings are equally valid.

5. Scoring

gcc-formats score (or gcc_format_validators.score) reports three accuracies overall and per class, variety and digit script:

• exact: the prediction equals the canonical reference;
• exact_any: it equals the canonical reference or one of the listed variants;
• normalised: equal after normalisation itn-norm-v1 (digit scripts, thousands and decimal separators, Arabic letter variants, diacritics, spacing inside numbers and codes).

As a floor, copying the input unchanged scores 0.000 normalised accuracy for ITN and 0.000 for TN on the test split (qa/reference_identity_scores.json).

Baseline. No system baseline is reported in this version. NVIDIA NeMo-text-processing was tried as a baseline, but its pynini dependency has no macOS arm64 wheel and its source build needs the OpenFst C++ library, so it did not install cleanly on the machine used to build this release.

6. Format reference tables

country	format	length	check digit	evidence
SA	Saudi national ID number	10	luhn	secondary
SA	Saudi resident identity number (Iqama)	10	luhn	secondary
AE	Emirates ID number	15	luhn	secondary
EG	Egyptian national number	14	unpublished (not verified)	structure only
KW	Kuwaiti civil number	12	unpublished (not verified)	structure only
BH	Bahraini personal number (CPR)	9	unpublished (not verified)	structure only
SA	Saudi vehicle registration plate	3 letters + 1-4 digits	none	secondary

"secondary" means the rule comes from published secondary sources (listed per row in spec_sources) rather than an issuing authority; "structure only" means the length and embedded fields are checked but no check-digit rule is applied. Identity formats for Qatar, Oman and Jordan are not in this version because no citable public description of their structure was found. format_examples holds 147 synthetic examples (95 valid, 52 invalid with the reason) for testing validators.

• IBAN (iban_registry): 21 Arab League states with length, BBAN structure and bank/branch/account positions; 17 are in the SWIFT IBAN Registry and 4 (DZ, KM, MA, YE) come from schwifty's additional data. No IBAN structure is listed for SY. Every row has spec_confidence = "secondary": the structures are taken from the schwifty library (MIT), which derives them from the SWIFT IBAN Registry (ISO 13616).
• Phone plans (phone_plans): 22 Arab League states with calling code, trunk prefix, number lengths and the number-range patterns per type, from the libphonenumber metadata (Apache-2.0).
• Saudi plates (plates_sa): the 17 plate letters and their Latin equivalents: ا A · ب B · ح J · د D · ر R · س S · ص X · ط T · ع E · ق G · ك K · ل L · م Z · ن N · هـ H · و U · ى V.
• Egyptian governorate codes (eg_governorates): 28 codes used in digits 8-9 of the national number.

A valid result means well formed. It never means that a number has been issued, is in use or belongs to anyone; nothing in this dataset or the package looks anything up.

7. Hijri-Gregorian table (Umm al-Qura)

hijri_gregorian has one row per day from 1924-08-01 to 2077-11-16 (55,991 days; 1 Muharram 1343 to the end of Dhu al-Hijjah 1500), with weekday, Hijri date, Arabic and English month names, month length and Julian day number. hijri_month_starts has 1,896 months. The dates come from hijridate (MIT).

Every day was also converted with ICU's islamic-umalqura calendar (icu 78.2 unicode 17.0 node v22.23.2). They agree on 33,933 days (60.6%) and differ on 22,058; agreement is complete for 1420-1450 AH. 740 month starts differ. Every disagreement is listed in hijri_icu_disagreements. Counted by the number of days between the date ICU gives and the date hijridate gives for the same day, the disagreements are -2: 493, -1: 20,479, 1: 671, outside range: 415 ("outside range": ICU's date is not a valid hijridate date). The table keeps the hijridate values. In the hijridate data, 8 early months have 28 or 31 days (1343/9, 1345/5, 1345/8, 1348/11, 1348/12, 1349/10, 1349/11, 1364/8).

8. Number words and currency units

number_words has 8,404 rows: cardinals for 2,001 values (0-1000 and a seeded sample up to 10^12) in masculine and feminine counted-noun forms and in nominative and accusative/genitive case, and ordinals 1-100. Each nominative row also gives the CLDR spellout produced by ICU (words_cldr); after removing diacritics the two agree on 326 of 4,002 cardinal rows and 200 of 200 ordinal rows. The cldr_agrees column marks each row, so the differences can be reviewed one by one.

currency_units:

ISO	numeric	minor digits	units	25,500.50 in words
SAR	682	2	ريال / هللة	خمسة وعشرون ألفًا وخمسمائة ريال وخمسون هللةً
AED	784	2	درهم / فلس	خمسة وعشرون ألفًا وخمسمائة درهم وخمسون فلسًا
EGP	818	2	جنيه / قرش	خمسة وعشرون ألفًا وخمسمائة جنيه وخمسون قرشًا

Each row also has the dual, plural and accusative forms, the CLDR display names by plural category, the CLDR symbol and common Arabic abbreviations; SAR lists U+20C1 SAUDI RIYAL SIGN (Unicode 17.0).

9. Related work and overlap check

NVIDIA NeMo-text-processing (Apache-2.0) has Arabic test cases: 126 ITN and 118 TN pairs in the classes cardinal, decimal, fraction, measure, money, whitelist (commit ddadfb2a38d2). Candidate pairs that matched one of them, exactly or after normalisation, were drawn again; 0 pairs in this release match (qa/itn_tn_contamination.json). Format data builds on libphonenumber, schwifty, hijridate and Unicode CLDR, credited in LICENSE-sources.md and PROVENANCE.jsonl.

10. Limitations

• No native-speaker review yet (see section 4). Colloquial spelling varies widely; scoring with normalised and the variant lists reduces, but does not remove, spelling penalties.
• The MSA rules make choices where usage varies: مائة (with مئة as a variant), the CLDR plural category "other" for 101 and 102 (singular noun), and nominative case by default.
• Format rules marked "secondary" or "structure only" may be incomplete. Corrections are welcome.
• The day table follows hijridate; where ICU differs, the difference is listed rather than resolved.

11. Licensing

layer	licence
Test pairs, templates and compiled tables by Bayanat Labs	CC BY 4.0
Format facts (lengths, prefixes, field positions, check-digit rules, letter maps)	CC0 1.0
phone_plans (libphonenumber metadata)	Apache-2.0
IBAN structures (schwifty) and Hijri dates (hijridate)	MIT notices kept
words_cldr and cldr_* fields	Unicode License V3
Validator and scorer code on GitHub	Apache-2.0

Full texts: LICENSE (CC BY 4.0), LICENSE-APACHE-2.0 (for the phone_plans rows derived from libphonenumber) and LICENSE-sources.md (upstream notices).

12. Files

data/<config>/<split>-00000-of-00001.parquet, tables/*.csv, tables/formats.json, qa/ (cross-check summaries, overlap check, identity scores, stats.json), PROVENANCE.jsonl, MANIFEST.json, SHA256SUMS, CITATION.cff, CHANGELOG.md, LICENSE, LICENSE-APACHE-2.0, LICENSE-sources.md.

13. Citation

@misc{bayanatlabs2026gccformats,
  title  = {GCC Format Validators and Arabic ITN/TN Test Pairs},
  author = {{Bayanat Labs}},
  year   = {2026},
  note   = {Version 0.1.0},
  url    = {https://huggingface.co/datasets/Bayanat-Labs/gcc-format-validators}
}

14. Corrections and takedown

Write to hello@bayanatlabs.com with the config, row and the correction or concern.

---

بالعربية

The Arabic text has not yet been reviewed by a native Arabic writer; the English governs.

هذه الحزمة من بيانات لابز تضم 2,000 زوجًا اصطناعيًا بين الصيغة المنطوقة والصيغة المكتوبة في 10 فئات (الأعداد، والأعداد الترتيبية، والمبالغ، والتواريخ الهجرية والميلادية، والأوقات، وأرقام الهواتف، وأرقام الهوية، والآيبان، والنسب المئوية، والقياسات) بالفصحى واللهجة السعودية/الخليجية والمصرية، وبالأرقام الغربية والأرقام العربية الهندية.

وتضم أيضًا جداول مرجعية: صيغ أرقام الهوية مع مصدر كل قاعدة، وبنية الآيبان لـ21 دولة، وخطط ترقيم الهواتف لـ22 دولة، وحروف لوحات المركبات السعودية (17 حرفًا)، وجدول التقويم الهجري (أم القرى) يومًا بيوم (55,991 يومًا من 1343 إلى 1500 هـ) مع مقارنته بتقويم ICU، وكتابة الأعداد والمبالغ بالحروف (التفقيط) للريال السعودي والدرهم الإماراتي والجنيه المصري.

جميع أرقام الهوية والآيبان والهواتف قيم عشوائية تستوفي شروط الصيغة. Generated examples; not real personal data; any match to a real person is coincidental.

لم تتم مراجعة الأزواج من متحدثين أصليين بعد، والقيمة في عمود review_status هي not_reviewed_by_native_speaker.

للتصحيحات وطلبات الإزالة: hello@bayanatlabs.com