Bayanat-Labs/arabic-wikipedia-articles-sections

Arabic and Egyptian Wikipedia Articles and Sections

2,967,774 current-revision article records and 14,240,743 parser-derived sections, from the Arabic and Egyptian Arabic Wikipedia editions' official 1 October 2026 dumps. Each article includes source wikitext, nested heuristic sections, page and revision IDs, and article/revision/history links. No images are included. Wikipedia contributors created the underlying material; Bayanat Labs extracted and packaged this derivative.

Source edition	Article records	Derived sections	Source dump bytes	Source SHA-256
Arabic Wikipedia (arwiki)	1,334,057	5,786,684	2,023,850,490	3871abbd0240f7b2293e2a58d5911c97390333ba23428b49f1e6c5adb350d843
Egyptian Arabic Wikipedia (arzwiki)	1,633,717	8,454,059	299,281,321	99ceab5a4c02f103da22ea59c4a3dc64bdec4d379d2886021766f335177d54f9
Total	2,967,774	14,240,743

These source editions identify the originating wikis, not an independently verified language, dialect or naturalness label. In particular, Egyptian Wikipedia contains mass automated translations and templated or bot-generated pages; see Alshahrani et al. (2024). The corpus is not a human-curated dialect sample.

What the records contain

A record has wiki, page_id, revision_id, revision_timestamp, title, article_url, revision_url, history_url, exact raw_wikitext, its wikitext_sha256, source_dump_date, source_dump_sha256, and nested sections. Each section has section_index, heading, heading_level and extracted text. The original revision can be checked using revision_url; its contribution history is linked by history_url.

Only main-namespace, nonredirect pages with a nonempty current wikitext revision were retained. The train split is a loading convention, not a held-out evaluation partition; there is no guarantee of near-duplicate or cross-edition independence. The articles config combines both editions; arwiki and arzwiki select one edition each.

from datasets import load_dataset
repo = 'Bayanat-Labs/arabic-wikipedia-articles-sections'
articles = load_dataset(repo, 'articles', split='train', revision='v0.1.0', streaming=True)
row = next(iter(articles))
print(row['title'], row['revision_url'])
print(row['sections'][0]['text'] if row['sections'] else '')

Extraction and quality limits

Section text is generated mechanically by mwparserfromhell using flat heading/lead segmentation and markup stripping. Templates are not expanded, and this output does not reproduce the rendered article. It can lose references, tables, links, formulas, layout or textual content; it can also retain image/caption fragments, link syntax and other wikitext markup. It is not fully rendered clean article text. Raw wikitext is retained for alternative processing. Sections are not human-annotated, and this release contains no human assessment of extraction accuracy. Wikipedia articles may contain errors or bias.

A full-pass mechanical analysis produced the following observations. “Short” means the total Unicode code points across heuristic section texts are fewer than 200; it is not Wikipedia's stub designation or a quality judgment. The markup probe matches a limited list of literal wiki/ref/file/caption patterns; it can miss or falsely flag markup and is not an accuracy estimate.

Edition	Articles under 200 section-text code points	Articles with a markup-probe match	Empty/whitespace-only sections
arwiki	114,870	163,758	5,010
arzwiki	214,580	271,618	1,007

The independent audit compared every released source article and its raw wikitext, revision and provenance fields with the pinned dumps, and fully hashed the output shards. Section text was reproduced from wikitext on a deterministic hash-selected sample of 2,951 articles, not exhaustively or by human reviewers. Details, definitions and exact source hashes are in source and quality metadata and methods.

Rights and attribution

Wikipedia's contributor text is generally available under CC BY-SA, subject to the actual source revision and exceptions. Wikimedia's license-transition FAQ states that revisions predating the transition remain available under CC BY-SA 3.0; pages subsequently edited can be reused under CC BY-SA 4.0 under the Wikimedia terms. Check the revision, its history and applicable terms; a dump date alone does not establish one license version for every source revision. Retain attribution to Wikipedia contributors, the source page or revision links and license notice; indicate that the text was extracted and repackaged and follow applicable share-alike obligations.

Articles can incorporate third-party quotations or other content with separate permissions or fair-use conditions. The general Wikipedia text license does not establish clearance for every embedded item. The dataset contains no image files, although raw wikitext and extracted text may contain image references or captions. See DATA_LICENSE.md and ATTRIBUTION.md. Bayanat offers its own copyrightable extraction and packaging contributions, if any, under CC BY-SA 4.0 without changing underlying or third-party rights. Processing code has its own Apache-2.0 license and its third-party dependencies retain their own licenses.

Sources and citation

The official dumps and their Wikimedia byte sizes and SHA-1 checksums were verified before processing. Source URLs, SHA-1 and locally computed SHA-256 are included in metadata; the full released file hashes are in SHA256SUMS. The extraction code is in source.

@dataset{bayanat_arabic_egyptian_wikipedia_2026,
  title = {Arabic and Egyptian Wikipedia Articles and Sections},
  author = {{Bayanat Labs}},
  year = {2026},
  version = {v0.1.0},
  url = {https://huggingface.co/datasets/Bayanat-Labs/arabic-wikipedia-articles-sections}
}

Also credit Wikipedia contributors and the specific source articles and revisions when reusing their text.