Hugging Face's logo

Datasets:
SlayerLab
/
polish-dynaword

Tasks:
Text Generation
Languages:
Polish
Size:
1M<n<10M
ArXiv:
Tags:
polish
pretraining
dynaword
License:
Dataset card Files
xet
 Community
117
polish-dynaword / data /nonsa /nonsa.md
ppuzio's picture
ppuzio
nonsa: drop namespace 100 Cytaty and <poem> pages (12,304 -> 11,692 docs, v0.2.6 rights audit) (#73)
9bdf1f9
preview code
|
Raw History Blame Contribute Delete
5.98 kB

Nonsensopedia - Polish satirical wiki (content namespaces) - CC BY-SA 3.0

This slice is rebuilt deterministically from the pinned wikiteam3 dump nonsa.pl-20251009-history.xml.zst (https://archive.org/details/wiki-nonsa.pl-20251009, sha1-verified). Scope is the content namespaces: 0 (articles), 102 (NonNews), 104 (NonZrodla), 106 (Slownik), 108 (Gra), 110 (Forum), 114 (Poradnik); the latest revision of each page. Namespace 100 (Cytaty) and every page whose wikitext has a <poem> block are dropped (see "Rights gate"). Nonsensopedia is the Polish satirical encyclopedia - informal, humorous, creative Polish covering parody reference text, fake news, forum discussions and interactive games - a register absent from the corpus.

  • Non-redirect content pages enumerated: 26,781
  • Source slice retained after QA/dedup and the rights gate: 11,692 (15,089 rejected: 14,425 too little text after stripping, 27 non-Polish, 23 low letter ratio, 1 near duplicate, 1 too little text after redaction, 287 namespace 100 Cytaty, 325 <poem> block)
  • Proposed records: 11,692
  • Proposed tokens: 15,439,683 (cl100k_base proxy)
  • Characters: 40,567,986
  • Before the rights gate the shard had 12,304 records, 16,883,883 tokens and 44,279,318 characters; the other 11,692 records are unchanged.
  • Earlier upstream build (made before the rights gate, so it still contains namespace 100 and <poem> pages; this shard is not taken from it): https://huggingface.co/datasets/PiotrSty/nonsa-pl/tree/09f62dc3826159dbc8c0df7a95243a18ec52abda

License

CC BY-SA 3.0 per record, the site license of nonsa.pl. Where that is stated at the origin:

Attribution is per record via the canonical page URL (nonsa.attribution.jsonl); each page's history lists its contributors. The licence covers what the contributors wrote; it does not extend to third-party works quoted on a page (see "Rights gate").

Rights gate

Two rules, both in src/build_nonsa_pl.py (third_party_drop), applied after the QA and dedup gates so that no other page's decision changes. The dropped pages are listed with their namespace, title and reason in nonsa.dropped.jsonl (612 rows); the parquet carries neither.

  • Namespace 100 (Cytaty): 287 pages, 440,332 tokens. These pages are collections of quotations from third-party films, games, songs, performers and public figures. Dropped, since the site licence does not cover the quoted works.
  • <poem> block: 325 pages, 1,003,868 tokens. <poem> marks verse and lyrics, which may be third-party works. The tag is read from the raw wikitext in acquire (has_poem_block), because strip_wikitext removes it; a <poem> that is only mentioned inside an HTML comment, <nowiki> or <pre> does not count (two forum pages are kept for that reason). The tag is also used for dialogues, jokes, game logs and quoted forum posts, so the rule drops those pages too. That over-matching is deliberate: telling third-party verse from contributor verse and from plain formatting needs a per-page review. A page in both rules is counted under 100.

Not covered by the gate: quotations embedded in retained pages. 3,293 retained pages (28.2%) call a {{cytat}}, {{cytat2}} or {{cquote}} template (the non-nested calls hold about 0.80M wikitext characters, about 1.7% of the retained pages' wikitext), 23 use <blockquote>, and lyrics typed as plain text are not detected. These have not been reviewed.

Processing

Per-record page id, title, namespace, last-revision timestamp and canonical URL are preserved; page history is the contributor-attribution trail. Wikitext is stripped with mwparserfromhell (strip_wikitext in src/build_nonsa_pl.py). Before strip_code(), category links ([[Kategoria:…]]) are removed and file links ([[Plik:…]], Grafika:, File:, Image:) are replaced by their caption only (the last non-option parameter; thumb, 200px, right, alt=… are dropped); inline [[:Kategoria:…]] links keep their text. The same rule is re-applied to links left verbatim in the output because malformed markup (e.g. an unclosed '' in a caption) made the parser read them as text. Magic words (__NOTOC__ etc.) are removed. After this, 2 of 11,692 records still contain a NNpx| option string and 103 a literal {{ (templates with unbalanced braces in the source wikitext); both are left as-is. Lines occurring in >=1% of documents are removed as cross-document boilerplate; emails, labelled phones, IP literals, labelled national identifiers and 26-digit account candidates are redacted (pattern redaction is not complete anonymization); exact and near deduplication applied within source.

PII

Scrubber: in-build pattern redaction only (emails, labelled phones, IPv4/IPv6 literals, labelled PESEL/NIP/REGON, 26-digit account candidates; counts in nonsa.qa.json). SlayerLab/NERGAL has not been run. This source stays at release: None until the shard is re-run through NERGAL.

Limitations

Deliberately non-factual register (satire, parody of real Wikipedia/cultural material); quoted passages are not guaranteed disjoint from other corpus shards. Embedded {{cytat}}/<blockquote> quotations and plain-text lyrics are not screened (see "Rights gate"). Record model is a whole content page; individual entries are not split.