syamjithnk/arabic-corpus-audit

Arabic Corpus Integrity Audit

Author: Syamjith NK Date: 9 September 2026 Tool: arabic-lint 0.5.0

The question

The arabic_reshaper + python-bidi recipe converts Arabic to presentation forms and visual order. It is correct for renderers that do no shaping, and it silently corrupts text on ones that do. It appears in 3,168 indexed Python files on GitHub.

If that recipe ever ran upstream of a training corpus, every model trained or evaluated on it learned from text no Arabic reader would accept. Nobody had measured whether it did.

What is in here

One row per dataset, for 341 public Arabic datasets on the Hub.

field	meaning
dataset	Hub id
readable	whether the datasets-server could serve rows
reason	why not, when unreadable
rows_sampled, text_fields	sample size actually scanned
fields_with_findings	text fields containing stored presentation forms
pct	percentage of scanned fields with a finding
spans	number of distinct corrupted spans
severity	counts of stray / partial / reshaped spans
worst_severity	the highest severity present in the sample

Severity, and why it decides what the numbers mean

severity	forms in a span	what it indicates
stray	1	one glyph pasted from a PDF, or OCR residue
partial	2–4	a fragment, or a short pass through the recipe
reshaped	5+	reshape+bidi ran over the text before it was stored

Counting findings without this distinction is misleading, because the two have different causes and different fixes: a stray glyph is a character to correct, a reshaped run means the pipeline that wrote the file is the problem and every other file it touched needs checking.

Across the whole audit: 515 stray spans, 9 partial, 108 reshaped — and all 108 reshaped spans are in a single dataset.

datasets by worst severity	count
stray	16
partial	4
reshaped	1

Headline result

276 datasets readable · 26,318 rows · 119,517 text fields · 21 with any finding · 1 fully corrupted.

Yousefmd/arabic_ocr_dataset is 100% corrupted — 400 of 400 fields across four sample windows, 23 to 29 presentation forms per label, INITIAL/MEDIAL/FINAL/ISOLATED forms together and words in reversed order. That is the recipe's complete signature.

It matters because the corrupted field is the label of an OCR set: a model trained on it learns to emit presentation forms in visual order. This is the propagation path from a rendering bug into model weights, observed rather than argued for.

⚠️ Its scale, stated plainly: 17 downloads, 0 likes. This proves the mechanism reaches training data. It is not evidence that widely-used Arabic corpora are affected.

The other 20 flagged datasets had no reshaped runs at all — 515 stray and 9 partial spans, most often U+FE91 BEH INITIAL FORM, from OCR and copy-paste.

Measured, and it is worse than "different tokens". On tokenizers without a Unicode normalisation step, a word containing one stray presentation form is not split, it is replaced by the unknown token. Every character of meaning is discarded before the model sees it. One corrupted letter out of four is enough: بيت becomes [UNK].

tokenizer	effect of one stray presentation form
aubmindlab/bert-base-arabertv02 (Arabic-specific)	6/6 words → [UNK]
google-bert/bert-base-multilingual-cased	6/6 words → [UNK]
xlm-roberta-base (SentencePiece)	0/6 — identical tokens, unaffected
Qwen/Qwen2.5-0.5B (byte-level BPE)	word survives, but as entirely different tokens

The irony is that the two Arabic-capable BERTs destroy the word while the multilingual SentencePiece model is immune, because XLM-R normalises and they do not. Re-runnable: tokenizer_cost.py, tokenizer files only, no weights.

What this does not establish

• 65 datasets could not be read (server errors, no viewer, gated). The viewer is most often unavailable on the largest corpora, so this covers the mid and small end well and the head of the distribution poorly. No claim here reaches web-scale pretraining data.
• 100 rows per dataset cannot find a defect confined to a rare subset.
• Presence of a presentation form is evidence of stored glyph forms, not proof of the specific recipe. The distinction is why severity is reported separately.

A false positive worth knowing about

The first run reported 35.6% of an Islamic heritage OCR corpus as corrupted. Every hit was a false positive. Arabic Presentation Forms-A is interleaved: the ornate parentheses ﴾ ﴿ enclosing a Quranic quotation are U+FD3E/U+FD3F, inside the positional-form range, with honorific ligatures just above. Treating the block as uniform accuses correct heritage text.

Fixed in arabic-lint 0.5.0 by deriving the class from the Unicode character name rather than tabulated ranges. Anyone auditing Arabic text should check this before reporting a number.

Reproducing

pip install arabic-lint
python3 audit_all.py          # https://github.com/Syamjith-NK/arabic-lint

Citation

@misc{syamjith2026arabicaudit,
  title        = {Arabic Corpus Integrity Audit},
  author       = {Syamjith NK},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/syamjithnk/arabic-corpus-audit}}
}