Add plwiki talk pages (ns 1+5) - conversational Polish

#36
by PiotrSty - opened
Fabryka AI org

Add Polish Wikipedia talk pages (ns 1+5) - conversational register

Adds plwiki_talk: 43,054 records and 55,568,733 measured cl100k_base proxy tokens from the pinned official Wikimedia dump plwiki-20260901-pages-meta-current.

Source dataset: https://huggingface.co/datasets/PiotrSty/plwiki-talk-pages/tree/20c74a6da0cbaef5d2a644d0accf7e90262ea24c

Data sample

Rights and method

Same CC BY-SA 4.0 basis as the shipped wikipedia shard (Wikimedia Terms of Use section 7: https://foundation.wikimedia.org/w/index.php?title=Policy:Terms_of_Use/pl&oldid=584706), on a namespace complement: ns=1 (article talk) + ns=5 (project talk); ns=0 is covered by the existing wikipedia shard and user-talk ns=3 is excluded by scope decision (not measured). Dump shards sha1-verified against dumpstatus.json. Wikitext stripped via mwparserfromhell; >=1%-frequency boilerplate lines removed; email/labelled-phone/IPv4/parsed-IPv6/labelled-PESEL-NIP-REGON/26-digit-account-candidate patterns redacted (not complete anonymization); langid vote; exact + near dedup within source.

Overlap and Slayer ontology

The overlap audit compared source-prefixed record IDs only; zero collisions do not establish text novelty. Comments can quote article text, so quoted-fragment text dedup remains an integration gate. The manifest carries content-addressed Versions, Protocol, Run, Evidence, falsifiable Claims, Actors and typed lineage. Benchmark checks, abuse screening and controlled ablations remain pending. This PR proposes a source, not a merged or stable DynaWord release.

Fabryka AI org

@kacperwikiel Review requested.

Polish Wikipedia talk pages (ns 1+5) — conversational register

  • 43,054 records, 55.6M cl100k_base tokens, 144.6M characters
  • Source: pinned plwiki-20260901-pages-meta-current dump, 7 sha1-verified shards
  • Scope: namespaces 1 (article talk) + 5 (project talk); ns=0 covered by existing wikipedia shard, ns=3 excluded by scope decision (not measured)
  • License: CC BY-SA 4.0 per Wikimedia Terms of Use section 7; per-record page URL preserves contributor-history attribution

Processing

Wikitext stripped via mwparserfromhell; >=1%-frequency cross-document boilerplate lines removed; Unicode/whitespace normalization; PII pattern redaction (email, labelled phone, IPv4, parsed IPv6 including compressed notation, labelled PESEL/NIP/REGON, 26-digit account candidates — not complete anonymization); three-window langid vote; exact + near dedup within source (deterministic prefix-indexed capped-shingle Jaccard >= 0.90).

Rejection breakdown

Reason Count
too_little_text (<500 chars) 500,652
non_polish_text 677
too_little_text_after_redaction 39
low_letter_ratio 92
normalized_duplicate 36
near_duplicate 41

Before/after evidence

Three complete before/after samples are in data/plwiki_talk/plwiki_talk.sample.jsonl and the source repo's artifacts/sample.jsonl. They show prose, paragraph boundaries, signatures and disagreement preserved; whitespace normalized; IP patterns masked.

Overlap

The overlap audit compared source-prefixed record IDs only; zero collisions do not establish text novelty. Comments can quote article text, so quoted-fragment text dedup remains a target integration gate.

Limitations

Talk pages are informal: typos, edit disputes, occasional vandalism. Record model is a whole talk page; comments are not split. Benchmark contamination, abuse screening and controlled training ablation remain pending.

Source snapshot

Cannot merge
This branch has merge conflicts in the following files:
  • .gitattributes
  • src/sources.py

Sign up or log in to comment