# wikinews

Polish Wikinews

## Dataset description
- **Source (upstream):** https://dumps.wikimedia.org/plwikinews/20260601/
- **Domain:** news
- **Language:** Polish (pl)
- **License:** `CC-BY-2.5` or `CC-BY-4.0` per record (by last-revision date)
- **Created (range):** 2011-12-24, 2026-05-25 (date of the page's last revision in the dump)
- **Added:** 2026-06-15; rebuilt 2026-10-06 (UTC) from the official Wikimedia dump

## Licensing — traceable basis
Contributors' text on pl.wikinews.org is licensed CC BY: [pl.wikinews terms of use (CC BY 2.5 until 2024-12-16, CC BY 4.0 since)](https://pl.wikinews.org/w/index.php?title=Wikinews:Warunki_u%C5%BCycia&oldid=342796&action=raw) (read 2026-10-06 UTC). The site licence
(`rightsinfo`) reads "Creative Commons Attribution 4.0". Each row is labelled by its page's last revision in the dump:

| last revision (UTC) | `license` | documents | tokens | share of tokens |
|---|---|---:|---:|---:|
| before 2024-12-16 | `CC-BY-2.5` | 17,581 | 8,956,397 | 73.8% |
| 2024-12-16 or later | `CC-BY-4.0` | 6,805 | 3,184,958 | 26.2% |

pl.wikinews has its own terms page, which names CC BY 2.5 until 2024-12-16 and CC BY 4.0 from that
day. A page last edited on or after the cutoff can still hold
2.5 text from earlier edits, and the [CC BY 2.5](https://creativecommons.org/licenses/by/2.5/legalcode) legal code (§4(a)) has no clause that lets an
adaptation be licensed under a later version (BY-SA 3.0 §4(b)(ii) has one). The 4.0 label for such
pages rests on the site's terms page alone; whether that is enough is for the lawyer.
The terms page gives the day, not the time. The cutoff is midnight UTC, so the 6 pages last edited
on 2024-12-16 are labelled 4.0. The configuration change was merged 2024-12-16 08:07 UTC and deployed
by 2024-12-16 08:32:24 UTC ([Gerrit change](https://gerrit.wikimedia.org/r/changes/operations%2Fmediawiki-config~1101867), [T381421](https://phabricator.wikimedia.org/T381421);
the deployment time was read by the rights audit, not re-read), so 2 of those 6 were last edited before it.
The rule is `license_for_revision` in `src/rights.py`, tested at the cutoff second, across time
zones, and on a missing timestamp, which stops the build. v0.2.0-v0.2.5 labelled every row `CC-BY-2.5`.

The terms page puts articles created before 2005-09-25 in the public domain. The dump has only the date of the last
revision, so how many released articles were created then is not known; 0 (0 tokens) were last edited before it.
All rows carry a CC BY label; nothing is dropped or relabelled for this.

Evidence: `wikinews.license-evidence.json`.

## Attribution
Each row's `author` is `Wikinews contributors, <page URL>`, e.g.
`Wikinews contributors, https://pl.wikinews.org/wiki/2005-03-18:_Barroso_w_Polsce`. The page history names the contributors; the terms page asks for a link to the source article, which the URL gives.
`wikinews.attribution.jsonl` has one line per row: `id`, `pageid`, `revid` (the revision
in the dump), `title`, `url`, `last_revision_at`, `license`. The row `id` is
`wikinews_<page id>` and `created` is the date of `last_revision_at`.

The text is modified: wiki markup is converted to plain text (see Provenance).

## Provenance
Parsed from the official Wikimedia dump `plwikinews-20260601-pages-articles` by
`src/fetch_wikimedia.py`; the file (21,980,828 bytes, sha1 `bfad5f33667756de4ded97756d370d7fdec39bcb`) was matched against the
dump's `dumpstatus.json` and `sha1sums.txt` (`wikinews.build.json`). v0.2.0-v0.2.5 used the same dump
date; the text set is the same (the rebuild adds the last-revision time).

Articles (namespace 0) only. Redirects, empty pages and pages shorter than 200 characters are skipped.
Wikitext is converted with mwparserfromhell `strip_code(normalize=True, collapse=True)`, as in v0.2.5.
Templates and comments are dropped. `<ref>` bodies, category links (`Kategoria:...`) and file-link
options (`thumb|...`) are not removed and stay in the text; the wikipedia fetcher's `strip_article`
removes them, this one does not, because that would change the text set. Of two pages with the same
text (sha1 of the stripped text), the lower page id is kept.

65,253 pages scanned → 47,845 in namespace 0 → 23,366 redirects, 0 empty pages, 0 parser failures, 87 shorter than 200 characters skipped, no exact duplicates → 24,392 read by the build → 24,386 released.

## Statistics
| documents | characters | tokens (tiktoken proxy) |
|---:|---:|---:|
| 24,386 | 32,524,646 | 12,141,355 |

The same text set as v0.2.5: documents, characters and tokens are unchanged; only labels, ids, authors and dates differ.

## Per-document license metadata
| license | documents |
|---|---:|
| `CC-BY-2.5` | 17,581 |
| `CC-BY-4.0` | 6,805 |

Author metadata present for **24,386** documents.

## Filters applied (build_dynaword.py)
- drop documents < 200 chars: **0** (the fetcher already dropped 87)
- drop non-Polish (diacritic ratio): **6**
- exact cross-source dedup (sha1): **0**
- OCR alpha-ratio < 0.70 (OCR sources only): **0**
- read 24,392 → kept 24,386

Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical
Llama-3 count is computed at release.

## Not checked
- pastes from copyrighted sources, including articles copied from press agencies or other publishers (no sample was run)
- the 2.5 → 4.0 label of pages edited after the cutoff (a reading of the site's terms page, not a lawyer's opinion;
  the lawyer's answer to E1, the wiki 3.0/4.0 labelling, is still pending, and whether it extends to 2.5 → 4.0 is not settled),
  and the time of day of the switch
- how many released articles were created before 2005-09-25, the public-domain date in the terms page
- text imported from elsewhere under another licence or with extra attribution requirements
- quotations inside articles (Art. 29 of the Polish copyright act)
- personal data (the NERGAL scrub runs before release)