Bayanat-Labs/arabic-ner-offset-fixtures

Arabic NER Offset Fixtures

By Bayanat Labs. 200 synthetic test cases for checking Arabic named-entity annotations: 100 dev cases and 100 acceptance cases across 18 categories of offset and tagging defects, plus valid records. Each case is a small annotation file in one of 6 formats with the settings needed to read it and whether a correct validator should accept it. They are the test cases of the Arabic NER Offset Validator, version 0.1.0.

Splits

Split	Cases	Valid	Invalid	How it was written
dev	100	25	75	Before the validator code; each case states the exact codes and repair suggestions expected.
acceptance	100	26	74	Separately, from the category list only, without the validator code or the dev cases; each case states its category, whether it is valid, and the defect in one sentence.

The splits use different sentences. Not a training set: no model benchmark is derived from it.

Categories

	Category	Code	Dev cases	Acceptance cases
a	Astral character (emoji) before the span: UTF-16 vs code point offsets	E014, E015	5	5
b	UTF-8 byte offsets	E014, E015	5	5
c	NFC composition drift	E011, E004	5	5
d	NFKC ligature expansion	E011	5	5
e	Span edge between a letter and its haraka or shadda	E004	5	5
f	Tatweel removed after annotation	E011	5	5
g	Bidi or zero-width control at a span edge	E012	5	5
h	Clitic boundary under the include or exclude policy	E008	6	7
i	Span edge inside a word or token	E007	5	5
j	Overlapping spans	E005	5	5
k	Nested spans under the forbid or allow policy	E006	5	5
l	Empty, zero-length or reversed span	E002	5	5
m	Invalid tag or tag transition (IOB2, BIOES, BILUO)	E009, E010	7	6
n	CRLF vs LF line-break shift	E013	5	5
o	UTF-8 byte order mark	E018	5	5
p	Span text differs from text[start:end]	E003	5	5
q	Arabic punctuation at a span edge	W102	5	5
r	Latin-script entity inside right-to-left text	E016, E012, E007, E008, E003	5	5
v	Valid records that must produce no error or warning	none	7	7

Fields

Field	Meaning
fixture_id	dev-<category>-NN or acc-<category>NN
split	dev or acceptance
category, category_name	category letter (a to r, v) and name
valid	true when a correct validator reports no error
format	canonical, doccano, label_studio, spacy, conll or hf_tokens
options	settings the format cannot state (JSON): offset_unit, scheme, nested_policy, clitic_policy, label_set, tag_names
input_payload	the file content, as a string (a byte order mark and CRLF line breaks are kept)
expected_error_codes	dev only: the exact error and warning codes expected (JSON list)
expected_suggested_export	dev only: the expected repair suggestions (JSON)
expected_problem, author_note	acceptance only: the defect in one sentence, and how the case was built
author, reviewed_by, synthetic, license	authorship; no case has been reviewed by a native Arabic writer (reviewed_by is empty); all cases are synthetic; CC BY 4.0

data/*.jsonl hold the same records as data/*.parquet. MANIFEST.json gives counts and SHA-256 checksums, SHA256SUMS the file checksums.

Use

from datasets import load_dataset
ds = load_dataset("Bayanat-Labs/arabic-ner-offset-fixtures")

Run a case through the validator (pip install "git+https://github.com/Bayanat-Labs/arabic-ner-validator@v0.1.0"):

import json
from arabic_ner_validator import validate
case = ds["dev"][0]
report, _, _ = validate(case["input_payload"].encode("utf-8"), dict(json.loads(case["options"]), format=case["format"]))

Results of the validator, version 0.1.0

Dev: 100 of 100 exact. Acceptance: 96 of 100; the four that do not pass are explained in the validator's docs/QA.md.

Sentences

Original Modern Standard Arabic, conversational and Gulf sentences written for these sets. Names are places, organisations, public or historical figures, or invented names; no sentence describes a real private person. The Arabic text has not yet been reviewed by a native Arabic writer; the English governs.

Licence and citation

CC BY 4.0. Bayanat Labs (2026). Arabic NER Offset Fixtures, version 0.1.0. https://huggingface.co/datasets/Bayanat-Labs/arabic-ner-offset-fixtures

Questions and corrections: hello@bayanatlabs.com