README.md

metadata

pretty_name: Wikipedia EN-FA Accessibility Bridge
license: cc-by-sa-4.0
language:
  - en
  - fa
task_categories:
  - translation
  - text-generation
tags:
  - wikipedia
  - persian
  - farsi
  - accessibility
  - parallel-corpus
configs:
  - config_name: articles_en
    data_files:
      - split: train
        path: data/articles_en/*.parquet
  - config_name: articles_fa
    data_files:
      - split: train
        path: data/articles_fa/*.parquet
  - config_name: parity_index
    data_files:
      - split: train
        path: data/parity_index/*.parquet
  - config_name: full_articles_en_without_fa
    data_files:
      - split: train
        path: data/full_articles_en_without_fa/*.parquet
  - config_name: full_articles_fa_without_en
    data_files:
      - split: train
        path: data/full_articles_fa_without_en/*.parquet
  - config_name: full_parity_index
    data_files:
      - split: train
        path: data/full_parity_index/*.parquet

Wikipedia EN-FA Accessibility Bridge

Current, attributable English and Persian Wikipedia article snapshots for pages created during a 69-day recency window, plus an EN↔FA counterpart index and static accessibility signals.

The reproducible full baseline is the official 2026-08-01 Wikimedia dump: 6,289,549 English articles without Persian, 129,821 Persian articles without English, and 22,277,907 namespace-0 pages in the combined parity index. Redirects are retained in the parity index but excluded from translation-ready article configs. This first release covers 2026-06-03T00:00:00Z through 2026-08-11T20:23:26Z. It contains 133,776 current English pages and 12,716 current Persian pages. Among the successfully fetched pages, 133,423 English pages lack a Persian counterpart and 5,868 Persian pages lack an English counterpart. Full translation- ready source text is included for 37,592 recently created English gaps and 1,132 recently created Persian gaps; pages that already have a counterpart remain in the compact parity index. The larger touched-page change ledger is retained by the builder but is outside this new-entry release.

Configurations

• articles_en: English wikitext and readable plain text for recently created non-redirect articles that lack a Persian counterpart.
• articles_fa: Persian wikitext and readable plain text for recently created non-redirect articles that lack an English counterpart.
• parity_index: compact status and EN↔FA counterpart map for all pages created in the window, including pages deleted or unavailable by fetch time.
• full_articles_en_without_fa: full English snapshot gaps as of 2026-08-01.
• full_articles_fa_without_en: full Persian snapshot gaps as of 2026-08-01.
• full_parity_index: every namespace-0 page in both 2026-08-01 snapshots.

Every article row retains page ID, current revision ID/timestamp/SHA-1, page URL, history URL, and Wikidata QID when available. Recent article rows additionally retain first and last event timestamps in the source window and whether the page was created during that window; full-snapshot rows retain dump date and shard.

What “recent” means

Wikimedia's monthly MediaWiki History dumps are used only as a metadata index to discover namespace-0 page IDs created from June 3 through July 31. The live MediaWiki Action API supplies the actual current wikitext, plain-text extract, revision provenance, Wikidata item, and language link. RecentChanges supplies the August portion. Therefore this is a dataset of current snapshots of recently created pages, not a dump of every historical revision body.

Accessibility fields

The article configs include counts and flags for heading-level skips, images without explicit wikitext alt=, tables without detected header markup, math, references, and mixed Persian/Latin directional text. These are triage signals, not a claim of WCAG conformance or a substitute for testing rendered pages with screen readers. A later semantic-HTML pass can prioritize pages flagged here.

Privacy and safety

The source history contains contributor metadata, but this pipeline deliberately does not retain editor names, IP addresses, user IDs, group membership, edit comments, or blocks. It stores page/revision provenance needed for attribution and reproducibility only.

Licensing and attribution

Wikipedia article text is available under the Creative Commons Attribution- ShareAlike 4.0 license, with additional terms described by Wikimedia. Preserve the page URL, history URL, and revision fields when redistributing or creating translations. MediaWiki History metadata and Wikidata structured data have CC0 terms; because this dataset includes Wikipedia text, the repository is labeled CC BY-SA 4.0.

Sources:

• https://wikitech.wikimedia.org/wiki/Data_Platform/Data_Lake/Edits/MediaWiki_history_dumps
• https://www.mediawiki.org/wiki/API:Recentchanges
• https://www.mediawiki.org/wiki/API:Revisions
• https://foundation.wikimedia.org/wiki/Policy:Terms_of_Use

Translation workflow

This repository is the canonical source layer. Machine translations should go in a separate derived dataset carrying the source page/revision identifiers, teacher model and exact checkpoint, prompt, decoding settings, automated quality signals, and human-review status. Generated translations are not silently mixed into the Wikipedia source rows.

Reproduction

The resumable builder and tests live in the Motarjem project. The release manifest records every Parquet shard's row count, byte size, and SHA-256 digest.