Bayanat-Labs/arabic-normalization-conformance

Bayanat Labs Arabic Normalization Conformance Fixtures · حالات اختبار المطابقة لتوحيد النص العربي

80 test cases (dev 64, acceptance 16) that pin down exactly what a named Arabic normalisation policy must output: the text, the rules applied, the lossy flags and the offset map. Version v0.1 (2026-10-15) · Library: arabic-normalize 0.1.0 · Demo: Arabic Text Normalizer · Arabic card: README.ar.md Published by Bayanat Labs (bayanatlabs.com), which is not affiliated with Space42 or the ARBML bayanat library.

بالعربية: 80 حالة اختبار كتبتها بيانات لابز (حالات التطوير 64، وحالات القبول 16)، تحدد بدقة ما يجب أن تخرجه سياسة مسمّاة لتوحيد النص العربي: النص الناتج، والقواعد المطبّقة، وتنبيهات فقد المعلومات، وخريطة الإزاحات. البطاقة الكاملة بالعربية في README.ar.md.

1. What is it

Arabic text normalisation (removing harakat and tatweel, folding alef variants, mapping presentation forms, handling zero-width characters) is done differently by every tool, and papers rarely say exactly what they did. These fixtures turn three named, hash-identified policies of the open-source library arabic-normalize into testable cases, so that any implementation, in any language, can check that it gives the same output. They are the conformance suite of arabic-normalize; the Python package and its JavaScript port both pass all of them.

Intended uses: testing an Arabic normaliser against a named policy; porting arabic-normalize to another language; documenting the edge cases of Arabic Unicode (presentation forms, combining hamza, ZWNJ, emoji ZWJ sequences, bidi controls, lone surrogates). Out-of-scope uses: training or evaluating language models (the set is tiny and synthetic); linguistic statistics about Arabic; any claim about how often a phenomenon occurs in real text.

2. What is being released

Config	Split	Content	Licence	Records
dev	test	development fixtures across 17 families	CC BY 4.0	64
acceptance	test	held-back fixtures with code-point families and rule combinations not in dev	CC BY 4.0	16

The same rows are in jsonl/dev.jsonl and jsonl/acceptance.jsonl. Every record was written by Bayanat Labs for this suite (synthetic: true); none is taken from a corpus. The 20 rejected-input cases (invalid UTF-8, over-cap requests, invalid policies) stay in the GitHub repository (tests/fixtures/rejected/), because they are not text-to-text cases.

3. Changelog

Version	Date	Change
v0.1	2026-10-15	First release: 80 fixtures (dev 64, acceptance 16) from arabic-normalize 0.1.0
v0.1	2026-10-15	Card revision: wording, contact address, product name and review status, and the English rationale of dev-021; no input or expected value changed

4. How to load

from datasets import load_dataset
dev = load_dataset("Bayanat-Labs/arabic-normalization-conformance", "dev", split="test")
acc = load_dataset("Bayanat-Labs/arabic-normalization-conformance", "acceptance", split="test")

# check an implementation
from arabic_normalize import normalize_detailed   # or your own normaliser
for fx in dev:
    # exact input, lone surrogates included, rebuilt from the code points
    text = "".join(chr(int(c[2:], 16)) for c in fx["input_codepoints"].split())
    r = normalize_detailed(text, fx["policy_id"])
    assert r.text_norm == fx["expected_output"], fx["fixture_id"]

5. Breakdown

Policy	SHA-256	Fixtures
asr-eval-v1	5d866146536a23b90c0882044dae5f55f2484f0635e206828246eabeb69668de	15
bayanat-norm-v1	ebdb82000d0762a65120ffa9311bc6f01b481b19d4348ea5bedb3ab23d759321	52
raw-safe-v1	57c3d9f03fb9b6854f2527cd1e8a47cadc266f47f02dda94652306dacea03bb5	13

Family	Fixtures
alef_hamza	7
alef_maqsura	4
arabic_compat_outside_pf	1
arabic_extended_letters	1
bidi_bom	4
bidi_isolates	1
c1_controls	1
decomposed_residue	1
diacritics	6
digits	4
digits_separators	1
emoji_zwj	2
final_nfc	1
format_controls	1
hamza_carriers	2
honorific_signs_extended_marks	1
ligatures_honorific	3
mixed_latin	3
nbsp_ornate	1
nfc_compose_reorder	1
persian_urdu	3
pf_marks	2
presentation_forms	5
punctuation	4
quranic_superscript	4
rejected_span	1
surrogates_nonchars	1
taa_marbuta	4
tatweel	3
whitespace	2
zero_width	5

• Fixtures whose expected output is lossy (non-empty expected_lossy_flags): 24. Fixtures with an informational flag: 7. Fixtures with removed invalid code points (expected_rejected_spans): 3.
• Fixtures that assert the full offset map: 11; that assert the diacritization ratio: 10. The reference_offset_map column gives the offset map for every fixture.
• Rule ids covered by expected_rules: 22 (alef_maqsura.to_yeh, alef_variants.map, bidi_zw.bidi_control, bidi_zw.zwj, bidi_zw.zwnj, bidi_zw.zwsp, canonical_order.final, canonical_order.nfc, diacritics.strip, digits.to_ascii, presentation_forms.map, punctuation.strip, quranic_marks.strip, reject_invalid.control, reject_invalid.lone_surrogate, reject_invalid.noncharacter, strip_controls.bom, strip_controls.format, superscript_alef.strip, taa_marbuta.to_heh, tatweel.strip, whitespace.collapse).
• Families that occur only in acceptance: arabic_compat_outside_pf, arabic_extended_letters, bidi_isolates, c1_controls, decomposed_residue, digits_separators, emoji_zwj, final_nfc, format_controls, honorific_signs_extended_marks, nbsp_ornate, nfc_compose_reorder, pf_marks, surrogates_nonchars.
• Input code points: 602; expected output code points: 517; reference change-log entries: 158; fixtures whose text the policy leaves unchanged: 20.

All numbers above are computed by scripts/build_conformance_dataset.py (GitHub repository) from the released files.

6. Baseline

Implementation	Fixtures passing every assertion
arabic-normalize 0.1.0 (Python)	80 of 80
@bayanat-labs/arabic-normalize 0.1.0 (JavaScript)	80 of 80

The JavaScript port also gives output identical to the Python package in 560 of 560 reference cases (these fixtures under 7 policies), including the offset map and change log (js/conformance.json in the GitHub repository).

7. Structure

Field	Type	Description
fixture_id	string	Stable id (dev-001, acc-001, ...)
split, family	string	Source split and the phenomenon the case tests
policy_id, policy_sha256	string	Policy the case is written for, and its SHA-256 (canonical policy JSON without meta)
input	string, nullable	Input text (may contain invisible characters and noncharacters). Null when the input holds a lone surrogate, which UTF-8 cannot store: rebuild it from input_codepoints
input_has_lone_surrogate	bool	True for those inputs (acc-010)
input_escaped, input_codepoints	string	The same input as Python-style escapes (\uXXXX, \UXXXXXXXX) and as space-separated U+XXXX code points (always exact)
expected_output, expected_output_codepoints	string	Required output
expected_rules	list of string	Sorted unique rule ids that must appear in the change log
expected_lossy_flags, expected_flags	list of string	Required lossy and informational flags
expected_offset_map	list of [int, int], nullable	Required offset map (output code point to input span); null when not asserted
expected_rejected_spans	list of {start, end, reason}	Invalid code points that must be removed and reported
expected_diacritization_ratio	float, nullable	Required ratio; null when not asserted
reference_offset_map, reference_changes	list	Full offset map and change log of arabic-normalize 0.1.0 (informative; computed, not authored). Each change has before/after text (null for a lone surrogate) and before_codepoints/after_codepoints
reference_library, unicode_version	string	arabic-normalize 0.1.0, 18.0.0
rationale_en, rationale_ar	string	Why the output is what it is
author, created, label, synthetic	string, bool	Authorship and the synthetic/illustrative label
reviewed_by, review_verdict	string, nullable	Human native review; null until signed

Offsets are code-point indices into input, zero-based and end-exclusive (not UTF-8 bytes, not UTF-16 units).

8. Creation

• Written in scripts/author_fixtures.py of https://github.com/Bayanat-Labs/arabic-normalize (source commit recorded in MANIFEST.json). The expected values were written before the engine existed, and the authoring script never calls the engine. The expected_* fields are the authored expectations; reference_* fields were added by scripts/build_conformance_dataset.py, which also checks that the reference output meets every expectation.
• Unicode data: UCD 18.0.0. Specification: docs/SPEC.md (arabic-normalize-spec-1).
• No web collection, no crawler, no third-party text.

9. Quality and human review

Every fixture passes in both implementations (section 6). A review of every lossy fixture and of the policy cards, made separately from their writing, is in docs/LOSSY_REVIEW.md in the GitHub repository (15 Oct 2026); it is not a native-speaker review. A named human native linguist has not yet signed tests/fixtures/REVIEW.md; until then reviewed_by and review_verdict are null. The Arabic text has not yet been reviewed by a native Arabic writer; the English governs.

10. Considerations

• Personal data: none. Common first names (for example مصطفى) appear only as spelling examples.
• Limitations: the pilot set is small (80 cases) and covers three policies; it shows behaviour, not frequency. Persian and Urdu words appear only to test that their letters are kept.
• Sensitive content: none known. Some fixtures contain Quranic-style spelling and the honorific ligature ﷺ to test how marks and ligatures are handled.

11. Licensing

Right	Status
Use in software tests, documentation and research	Allowed with attribution (CC BY 4.0)
Model training	Allowed with attribution; not a training set
Redistribution and modification	Allowed with attribution

Attribute as: "Arabic Normalization Conformance Fixtures, Bayanat Labs, CC BY 4.0". The arabic-normalize code is Apache-2.0; this dataset is data and is CC BY 4.0.

12. Citation

@misc{bayanatlabs_arabic_normalization_conformance_2026,
  title  = {Bayanat Labs Arabic Normalization Conformance Fixtures},
  author = {{Bayanat Labs}},
  year   = {2026},
  note   = {Version v0.1, arabic-normalize 0.1.0},
  url    = {https://huggingface.co/datasets/Bayanat-Labs/arabic-normalization-conformance}
}

Also see CITATION.cff. To cite a normalisation, name the policy id and its SHA-256.

13. Provenance

Source	URL	Licence	Basis	Date
Bayanat Labs, authored for this suite	https://github.com/Bayanat-Labs/arabic-normalize	CC BY 4.0	own work	2026-10-15

Per-record ledger: PROVENANCE.parquet (fixture id, source commit, fixture SHA-256, licence). Checksums: SHA256SUMS; file list: MANIFEST.json.

14. Claims

No comparative claims. See CLAIMS.md.

15. Takedown and contact

Email hello@bayanatlabs.com. We act on removal requests within 30 days and on platform requests within 24 hours.

Need production-grade, rights-cleared data?

This free release covers normalisation test cases. For licensed Arabic text and knowledge datasets with documented normalisation, see Bayanat Knowledge → https://www.bayanatlabs.com/products/knowledge