# Methods

The two official Wikimedia 20261001 current-revision `pages-articles.xml.bz2` dumps were pinned by exact size and Wikimedia SHA-1, then locally hashed with SHA-256. Names, URLs and hashes are listed in [metadata/source-and-quality.json](metadata/source-and-quality.json). The build streams bzip2 XML; it does not materialize the expanded XML. It selects only article namespace 0, nonredirect pages with a nonempty current wikitext revision. It preserves the raw text, revision timestamp and source links and computes the raw UTF-8 SHA-256.

The parser uses `mwparserfromhell` 0.7.2 to split flat headings and lead sections, then applies `strip_code(normalize=True, collapse=True)`. Templates are not expanded. The output is a heuristic representation, not a rendered page: removed content, malformed text, image or caption fragments and other markup can remain. Section headings and extracted text were not evaluated by humans.

The independent audit re-read the exact verified dumps and checked every released article's source text, revision and provenance fields, plus the output shard bytes. It reconstructed section contents for a deterministic sample selected by SHA-256 of the wiki and page ID modulo 1024, not for every article. A separate full-pass quality analysis reports mechanical text-length, empty-section and limited markup-pattern observations; these do not establish language, naturalness, article quality or accuracy. Counts and definitions are in [metadata/source-and-quality.json](metadata/source-and-quality.json).

The `train` split is an ingestion convention, and this is not a benchmark with held-out groups, near-duplicate exclusion or semantic relevance labels. No language or dialect is inferred from source edition name. External quotations or fair-use content can remain in the source wikitext, requiring separate rights assessment.