Hatim2221/arabic-wikipedia-clean

Arabic Wikipedia (Cleaned)

A cleaned, deduplicated extraction of Arabic Wikipedia articles, built from a Kiwix ZIM dump for use in small-scale language model pretraining.

Dataset Details

• Source: Arabic Wikipedia (wikipedia_ar_all_mini), via a Kiwix .zim archive
• Language: Arabic (ar)
• Format: Parquet, with train and validation splits
• Fields:
  • id (string): stable identifier derived from the article's source path
  • title (string): article title
  • url (string): the article's path within the source ZIM
  • text (string): cleaned article text (title prepended, see below)

Cleaning pipeline

Raw HTML was extracted from the ZIM archive and processed to:

• Strip navigation, infoboxes, reference/citation blocks, and other non-prose HTML
• Remove the Arabic Wikipedia license-footer boilerplate
• Strip leading calendar-grid noise from day-of-year pages (structural detection, not a hardcoded date list)
• Drop disambiguation ("توضيح") pages
• Filter articles below a minimum length threshold
• Deduplicate exact-duplicate text
• Cap near-identical templated stub articles (e.g. mass-produced asteroid entries) to at most a few examples per structural template, to reduce training-data skew

See the accompanying extract_zim.py script for the full, exact logic.

License and Attribution

This dataset derives from Arabic Wikipedia and is distributed under the same terms: CC BY-SA 3.0, and also available under the GFDL. Text content originates from Wikipedia contributors; see https://ar.wikipedia.org/wiki/ويكيبيديا:حقوق_النشر for full license details.

If you redistribute or build derivative datasets/models from this data, you must comply with the share-alike and attribution terms of CC BY-SA 3.0.

Known Limitations

• Length distribution is heavily right-skewed (most articles are short stub entries), typical of a full Wikipedia dump -- see the corpus analysis for details before using this as a sole pretraining source.
• Some short factual templates (e.g. astronomical object entries) remain in the data at a capped rate; they were not fully removed, only downsampled.
• No PII-specific filtering was applied beyond what Wikipedia's own content policies already enforce.

Citation

If you use this dataset, please cite Wikipedia itself as the underlying source, e.g.:

Wikipedia contributors. "Arabic Wikipedia." Wikipedia, The Free Encyclopedia.
https://ar.wikipedia.org