Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 87

What and why

Relabels wikibooks (Polish Wikibooks) per record: CC-BY-SA-3.0 or CC-BY-SA-4.0 by the page's last revision. v0.2.0-v0.2.5 labelled every row CC-BY-SA-3.0. Part of the October 2026 rights audit, slayerlabs/polish-dynaword-v2#57 (artifacts/rights_audit_2026-10/reports/wikis_ugc.md, reports/citations.md L17-26).

Rebased onto main 262df1b (after #68 wikipedia, #77 wikinews and #71 eurlex merged), so it no longer carries #68's commits or the merge commits. Three commits: DROPPED_PAGES shared code (src/rights.py, src/fetch_wikimedia.py, src/test_wikimedia_contract.py; rebased from b94771e), the wikibooks labels (from 15c123b) and the 46 withheld pages (from c76132b). Every file is identical to the pre-rebase head c76132b except src/sources.py and artifacts/source_findings.md, where main's other sources are kept and the wikibooks changes are the same. The lawyer's answer to E1 (wiki 3.0/4.0 labelling, slayerlabs/polish-dynaword-v2#56) is still pending; the labels apply the cutover dates of the rights-audit task (v2#57) and are not a legal opinion.

46 pages are withheld (maintainer approval 2026-10-06, relayed by the coordinator): 46 documents, 697,066 characters, 258,066 tokens (36 labelled 3.0, 10 labelled 4.0).

• wikibooks_23933 Historia powszechna/Rzeczpospolita pod panowaniem królów elekcyjnych: copied from sciaga.pl (follow-up check 2).
• wikibooks_27715 Od lamera do hakera/Linux: states its own CC BY-NC 3.0 terms, plus a translator's profit condition and an unattributed Wikipedia block (check 3).
• wikibooks_27602 Od lamera do hakera/Procesor: pasted external text, source link lost, no licence (check 3).
• the Mentalizm book, 43 pages: wikibooks_26962 (Mentalizm/Wprowadzenie, the notice allowing non-commercial use only, share-alike) and the 42 other pages including the hub wikibooks_26961, which carry no notice of their own and are withheld with the book (check 3).

They are listed with one reason each in DROPPED_PAGES["wikibooks"] (src/rights.py, same shape as COPIED_PAGES; the 42 chapters share one reason); fetch_wikimedia.py prepare skips them, stops if a listed id is not reached (not found, a redirect, not namespace 0, empty, short or a duplicate), and records them with their reasons in wikibooks.build.json. A page whose text is a copy of a withheld one is still dropped as a duplicate. The rest of "Od lamera do hakera" (hub 27691, 27584, 27716, 28327) stays released.

The v0.2.5 parquet has no last-revision timestamp, so the source is rebuilt from the official Wikimedia dump plwikibooks-20260601-pages-articles (sha1 198ddcbe… matched to the dump's sha1sums.txt) by src/fetch_wikimedia.py. Checked against the v0.2.5 parquet: same schema; the text multiset is v0.2.5's minus the 46 withheld pages (the first rebuild of this PR matched v0.2.5 exactly; the current parquet equals that rebuild minus the 46 rows, checked row by row); only license, id, author, created and added change. Row ids are now wikibooks_<page id> ("Przewodnik po instalacji Debiana 3.0": wikibooks_2419570 → wikibooks_1710).
author = Wikibooks contributors, <page URL> (v0.2.5: empty); created = the last-revision date (v0.2.5: one range string on every row); the parquet added column is 2026-07-19 (v0.2.5: 2026-06-15; the registry's added stays 2026-07-02, release stays 0.2.0). The shared code that moves license_for_revision from src/fetch_wikipedia.py into src/rights.py, rewrites src/fetch_wikimedia.py and adds src/test_wikimedia_contract.py is on main since #77; this PR adds DROPPED_PAGES on top of it.

Before → after

Rules, counts, hand-checks

src/fetch_wikimedia.py (acquire → prepare → build_dynaword.py → attribution → datasheet), then build_dynaword.py's own gates. Samples: seed 20261006 (random.Random), 8 per rule unless all were read.

Funnel: 34,967 → 12,838 in namespace 0 → 9,597 written → 9,066 kept. Labels before the language filter: CC-BY-SA-3.0 4,588, CC-BY-SA-4.0 5,009. Every row's label is rechecked against its last-revision time in the attribution sidecar by check_released (see Tests).

Label rule license_for_revision (src/rights.py): last revision before 2023-06-07T00:00:00Z → CC-BY-SA-3.0, at or after → CC-BY-SA-4.0; a missing or zone-less timestamp stops the build. Same rule and same midnight cutoff as the wikipedia PR. 2 pages were last edited on 2023-06-07 itself and get 4.0; the audit's 06-08 would put that day's 4.0 edits under 3.0, which 4.0 §3(b)(1) does not allow, and the dump has no switch time to cut more finely, so up to 2 of these may be 3.0 text under a 4.0 label.

Provenance and licence

• Wikimedia Terms of Use oldid 554823 (CC BY-SA 4.0 and GFDL, in effect June 7, 2023; sha256 c96ef93b…) and oldid 98524 (CC BY-SA 3.0, June 16, 2014; sha256 46617e35…); CC BY-SA 3.0 §4(b)(ii) (sha256 c33d7ebd…) and 4.0 §3(b)(1) (sha256 47c0a954…) legal code; pl.wikibooks.org rightsinfo CC BY-SA 4.0. Read 2026-10-06. A page edited after the switch can carry earlier 3.0 text under 4.0 via 3.0 §4(b)(ii) (inferred, not a lawyer's opinion).
• Attribution: each row's author links its page, whose history names the contributors; data/wikibooks/wikibooks.attribution.jsonl (LFS) has id, page id, revision id, title, URL, last-revision time and licence per row; wikibooks.license-evidence.json pins the sources above; wikibooks.build.json pins the dump and the counts.
• Registry: src/sources.py provenance, licence (license_spdx MIXED), traceable text and upstream updated; the card is generated by fetch_wikimedia.py datasheet (custom_datasheet); make_docs.py, README and CHANGELOG are not touched.
• Text conversion is the v0.2.5 one (plain mwparserfromhell.strip_code), so <ref> bodies, Kategoria: links and thumb| file options stay in the text. That is unchanged and stated in the card; fixing it would change the text of every release and is not part of this PR.

GFDL review (wikibooks)

The audit names 24 GFDL-mentioning documents (0.4% of tokens) and wikibooks_2419791, wikibooks_2419792. That count cannot be reproduced: its pattern is not recorded; the closest, \bFDL\b|Wolnej\s+Dokumentacji, gives 25 documents / 69,072 tokens and contains neither named document (they are now wikibooks_3153 / wikibooks_3155, the "Dive into Python" translation). A wider case-insensitive pattern gives 45 documents / 122,192 tokens (0.80% of the release; 0.78% before the withholding). All 45 were read and sorted (data/wikibooks/wikibooks.gfdl-review.json, one line of reason per document):

The 17 own-text and upstream documents (49,737 tokens) are the ones where a GFDL-only grant could matter; the largest is wikibooks_30538 ("Linux i Systemy Wbudowane", 35,619 tokens, "Udostępnione na licencji GNU Free Documentation License", last revision 2022-02-02, labelled 3.0). No label or drop follows from the review; whether to keep, relabel or drop those 17 is for the maintainers and the lawyer. Separately, 435 documents (259,065 tokens, 1.7%) were last edited before 2009-06-15, when the Wikimedia sites began dual licensing under CC BY-SA 3.0 and the GFDL; they carry CC-BY-SA-3.0, counted and not acted on. Two pages seen during the review and not checked: wikibooks_27716 (145 tokens) lists third-party course sources with their licences, among them CC-BY-NC 3.0 and "All Rights Reserved"; wikibooks_27071 (1,287 tokens) says it was moved in 2013 from the "Wolne podręczniki" portal with its creators' consent and does not state the original licence.

Dedup (reported, not fixed)

The text multiset is v0.2.5's minus the 46 withheld pages, so exact-text overlap with other sources can only have shrunk; it was not re-measured. No overlap_audit.json in the repo mentions this source.

Follow-up audit checks (2026-10-06)

Measured and reported first; the figures below are for the 9,112-document release before the withholding. The maintainer then approved withholding 46 pages (2026-10-06, see above): 23933, 27715, 27602 and the Mentalizm book. Nothing else was dropped or relabelled, and the recommendations below are as written before that decision.

1. The two documents seen in the GFDL review, read in full.

• wikibooks_27716 (Od lamera do hakera/Bibliografia, 145 tokens): a list of source names with the licence of each (CC-BY-NC 3.0, GNU FDL, GNU GPL, "All Rights Reserved"). It repeats no sentence of any of them; the NC and all-rights-reserved material is only named. Recommend keep.
• wikibooks_27071 (Fizyka dla gimnazjum, 1,287 tokens): one provenance sentence (moved in 2013 from Wolne podręczniki with the consent of the portal's creators, licence not stated) and a chapter outline; no third-party prose. Recommend keep, with the missing licence statement noted.
• Found by the keyword scan below: wikibooks_27715 (Od lamera do hakera/Linux, 3,761 tokens) states of its own text that it is under CC BY-NC 3.0 (translation by Paweł Niewiadomski of a work by Lars Wirzenius). That is a first-party non-commercial statement inside the text. Recommend drop, or a lawyer's decision. The book's hub page wikibooks_27691 warns that the licences of its texts are not uniform; all 6 released pages of the book are traced in section 3. wikibooks_26962 (Mentalizm/Wprowadzenie, 1,313 tokens): the author's notice allows redistribution and changes for non-commercial purposes only, share-alike. Lawyer question; drop if the answer is no.
• Keyword scan of all 9,112 documents: 23 contain non-commercial wording (56,308 tokens). Two state it for their own text (27715, 26962), one is the source list above (27716), and 20 are licence explanations, software lists, links, or say that an NC-licensed tool cannot be used here. 5 contain "all rights reserved": program banners in code listings (21436, 25400, 28622, 32057) and a licence explanation (24824); none is wiki text under that reservation.

2. Copyvio sample (audit item). Earwig Copyvio Detector API (copyvios.toolforge.org, public, no login), same method and size as the wikipedia agent's 80-page sample: 40 released pages per licence label (80), random.seed(20261006) then random.sample over the ids sorted by number within each label, action=search with use_engine=1, use_links=1, oldid = the dumped revision, 3 s between calls. Every hit at confidence 0.10 or more, and every one marked possible or suspected, was read by hand against the page history (first revision holding the shared passage) and the date of the source (page metadata or Wayback). Limits: Earwig finds pages a search engine returns; paywalled, blocked or unindexed sources are missed, and a later copy on the web does not show that the wiki text was original. 79 of 80 pages were checkable (wikibooks_4757 was deleted from the wiki on 2026-08-02 and Earwig answers bad_oldid).

Copied from a copyrighted source: 1 of 79. wikibooks_23933 Historia powszechna/Rzeczpospolita pod panowaniem królów elekcyjnych (10,753 tokens, label 4.0, page created 2011-03-28 by Tomekg.966): https://sciaga.pl/tekst/40230-41-wojny_polsko_tureckie_w_xvii_w_i_i_ii_polowa_przyczyny_skutki , confidence 0.964, 43.9% of the page's words shared, longest shared run 153 words. sciaga.pl has captures since 2007-10-11, so the source is older than the page; its licence was not examined. Recommend drop, or a lawyer's decision.

Hits read and cleared: wikibooks_56086 (public-domain text), 15616 (slideserve copy of 2014, later), 5702 (github, later), 56185 (gnuplot program output), 16383 (prezi created 2015-05-20; the passage entered the wiki on 2011-03-08 by Tomekg.966; later copy, upstream unresolved), 14012 (stock recipe phrase, sjp.pwn.pl).

Wilson 95% upper bounds: 1 of 79 is 6.8% (3.0: 0 of 40 gives 8.8%; 4.0: 1 of 39 gives 13.2%).

Extra: one contributor. 23933 and the passage of 16383 both come from Tomekg.966, who created 91 corpus pages in 2011 (54 Historia dla klas 7-8 szkoły podstawowej, 35 Historia powszechna, 1 Brudnopis, 1 Genealogia Piastów). The other 90 were run through the same Earwig call (not a random sample, so no new estimate for the source): 0 copied. Four hits at 0.10 or more were read: wikibooks_25499 (https://brainly.pl/zadanie/6982918 , 0.76; Brainly page created 2014-04-18, text on the wiki since 2011-11-29, later copy), 24465 (https://prezi.com/p/uk7qpes88rzs/poszukiwanie-morskiej-drogi-do-indii/ , 0.47; prezi created 2018-09-23, on the wiki since 2011-07-29, later copy), 24372 (https://www.historycy.org/index.php?showtopic=128688 , 0.60; the forum posts are from July 2015, on the wiki since 2011-05-13, later copy; the site refuses direct fetches, read through a Wayback capture), 25419 (http://hamlet.edu.pl/debeca-pasowanie , 0.68; shared passages are a medieval knight's oath, and the edit summary says the page pastes a Wikipedian's text). Wilson upper bound 0 of 90 is 4.1%; with 23933, 1 of 91 is 6.0%. For the three later copies the upstream of the 2011 text was not found.

3. Licence trace of the two books with non-commercial wording (2026-10-06, measured only). Method: full revision histories of all 49 released pages (MediaWiki API, prop=revisions, action=compare), the 49 dumped texts read against them, the Earwig run above on every page (all 49 answered), Wayback captures for the dead source site. No drop, no relabel.

Od lamera do hakera (hub wikibooks_27691): 6 released pages, 14,794 tokens, all labelled CC-BY-SA-3.0 (last revisions 2016 to 2021). The hub's outline has 2 more pages in the prefix (Karta Graficzna, Płyty kompaktowe); they are not in the dataset.

Totals: 14,794 tokens = 3,761 CC BY-NC 3.0 translation (27715) + 7,548 third-party text of unknown licence (27602) + 3,485 the contributors' own wiki text (27584, 27691, 27716, 28327).

27715 in detail. The wiki page was created 2013-07-22 ("new art", 10,373 bytes) from http://www.linuxpl.org/SAG/c1567.html (Wayback capture 2013-10-26 exists). The note naming CC BY-NC 3.0 is the wiki page's own; the source site's wording differs: the Wayback capture of linuxpl.org/SAG (2013-05-28) carries Wirzenius's notice (copying of the whole manual allowed with the copyright notice kept; modified versions and translations under the same terms) and a note by the translator asking for a share of any profit and to be told of commercial distribution. So the terms are a custom permission plus a translator's condition, not CC BY-NC as such. In addition a block of about 220 words ("POST") was added to the page on 2013-07-23; it is copied from pl.wikipedia "POST" (revision of 2013-03-12: 43 of 53 eight-word chunks found, longest run 174 words), which is CC BY-SA and not attributed here. Earwig: possible 0.641 against a prezi.com copy of that Wikipedia text.

Every page of the book was created by Kompowiec2 except 28327 (an IP); the hub warns, and the bibliography confirms, that the texts have different licences. The hub also links to 25 released pages that belong to other books (C, C++, CSS, Git, HTML, Java, Linux 387, Internet w praktyce, Komputer w praktyce and others); those are separate books and were not traced.

Mentalizm (hub wikibooks_26961): 43 released pages, 236,004 tokens (34 labelled CC-BY-SA-3.0 = 174,775 tokens, 9 labelled CC-BY-SA-4.0 = 61,229 tokens). Besides 26962 there are 42 pages, 234,691 tokens:

26961 (625, hub), 26963 (6,365), 26964 (14,766), 26968 (5,927), 26969 (453), 28768 (5,067), 28769 (1,970), 28770 (967), 28771 (2,665), 28772 (980), 28776 (2,154), 28777 (3,469), 28778 (960), 28779 (2,311), 28780 (9,148), 28781 (6,014), 28782 (6,092), 28784 (5,667), 28785 (1,130), 28789 (12,794), 28795 (5,603), 28796 (2,538), 30772 (934), 31385 (4,343), 31393 (6,547), 31432 (4,710), 31433 (3,754), 31434 (2,305), 31634 (9,255), 31635 (12,563), 31637 (12,104), 31649 (1,823), 31652 (3,177), 31653 (4,025), 31654 (4,387), 31656 (10,664), 31658 (18,760), 31662 (11,383), 31664 (6,539), 31667 (6,929), 31668 (8,914), 31669 (3,910).

• None of the 42 is the Wirzenius translation or any other third-party text found. None repeats the licence notice or contains non-commercial or all-rights-reserved wording about its own text (keyword scan; 28780 mentions in passing that GNU/Linux and "this book" are under liberal licences). The notice exists only on 26962.
• All 43 pages were created by one user, Pedros.lol, the author (empty user page). Other users' revisions are tiny: the largest are +247 bytes on the hub and +84 on 31653, the rest mostly under 10 bytes (typos, reverts, a Commons bot). On 28780, 28782, 28789 and 31635 some revision usernames are hidden.
• Earwig on the 42: nothing at 0.10 or more (highest 28789 0.091 goodreads, 31667 and 31668 0.065, 31658 0.057; all below the 0.10 reading threshold, none flagged). 0 of 42 copied, Wilson 95% upper bound 8.4%. Earwig is weak on translation and paraphrase; 31635 contains a few quoted paragraphs and examples from other publications, attributed and short.
• The notice on 26962 is inconsistent. Its text allows use for non-commercial purposes only, share-alike. The first revision (2013-01-02) had that wording; the author removed the non-commercial part on 2013-01-13 ("poprawka w licencji") and restored it on 2015-02-02 (rev 207241). The same day rev 207242 linked "Creative Commons" to the CC BY-SA 3.0 deed (creativecommons:by-sa/3.0/deed.pl), which has no NC. Four chapters (26963, 26964, 26968, 26969) were created on 2013-01-02/03 while the NC wording stood; most others from 2015-02 on. The edits were also submitted under the Wikimedia terms of use.
• The hub says the book is the basic version of the author's paid "Mentalizm - kompendium"; a 2016 talk-page comment calls the book an advertisement for it. Whether the NC wording is deliberate is for the lawyer to ask the author.

Recommendations as written before the decision (carried out: 27715, 27602, 23933 and the whole Mentalizm book, 43 pages, are withheld; the others stay): 27715 drop, or a lawyer's decision (the NC statement, the translator's profit condition and an unattributed Wikipedia block in one 3,761-token page); 27602 drop, or a lawyer's decision (7,548 tokens of an external text with a lost source and no licence); 27584, 27691, 27716, 28327 keep; 26962 lawyer; the other 42 Mentalizm pages (234,691 tokens) lawyer, because the notice sits on the book's introduction and the author's intent for the whole book is unclear, though no other page states NC.

No tool asked for a login or terms acceptance.

Open

• E1
• the 2 documents edited on 2023-06-07
• the 17 own-text / upstream GFDL documents (above)
• non-commercial and unlicensed text: wikibooks_27715, wikibooks_27602, wikibooks_26962 and the other 42 Mentalizm pages are withheld; nothing further is open on them here
• the 25 released pages that the "Od lamera do hakera" hub links in other books were not traced
• copyright pastes: wikibooks_23933 (copied from sciaga.pl) is withheld; 1 of 79 sampled, 0 of 90 other pages of the same contributor; of the 9,112 pages before the withholding, 8,942 were not checked
• the datasheet's "not checked" list still says "pastes from copyrighted sources (no sample was run)"; that is no longer true after the follow-up sample and is not changed here
• the markup residue

Tests

pytest -q src on main 262df1b with the wikibooks, wikinews and wikipedia data pulled: 486 passed. With this PR: 500 passed; the same 38 failures and 10 errors on both come from other sources' Parquets that were not pulled (LFS pointers). Before the rebase, on c76132b: 402 passed, 0 failed (361 on 15c123b, before the merges and the withholding). check_released("wikibooks", ..., dropped=(46, 697066, 258066)) pins counts equal to v0.2.5 minus the 46 pages' documents, characters and tokens (9,066 / 38,195,523 / 15,313,229; labels 4,425 and 4,641) and every row against the attribution sidecar, the evidence file, the registry entry and the committed card; test_wikibooks_withholds_three_copied_pages_and_the_mentalizm_book checks that exactly the 3 copied pages and 43 Mentalizm pages (with 26961 and 26962) are listed with a reason, are absent from the parquet and are recorded in wikibooks.build.json with the same reasons; the GFDL-review test is unchanged and passes on the regenerated wikibooks.gfdl-review.json (the review's 45 documents are unaffected, its shares are against the new total). The fixture tests of prepare now patch DROPPED_PAGES to a fixture list (empty by default) so they do not depend on the real one.

python3 src/render_sample.py data/wikibooks/wikibooks.parquet -n 3

id: wikibooks_61481
source: wikibooks
added: 2026-07-19
created: 2026-01-26
token_count: 232
license: CC-BY-SA-4.0
author: Wikibooks contributors, https://pl.wikibooks.org/wiki/Napary/Eleuterokok_kolczasty

Eleuterokok kolczasty 

 Eleuterokok kolczasty jest adaptogenem.
 Eleuterozydy zawarte w korzeniu pobudzają ośrodkowy układ nerwowy, zwiększają zdolności adaptacyjne organizmu, podnoszą wydolność psychiczną i fizyczną.
 Poprawia ogólne samopoczucie.

 Przeciwwskazania 

 Surowiec zielarski 
 Na napary wykorzystuje się wysuszone (całe lub rozdrobnione) korzenie eleuterokoka (Eleutherococci radix).

 Napary z eleuterokoku kolczastego 
 Herbatka z suszonego korzenia eleuterokoku kolczastego 
 Zalej 1-2 łyżeczki suszonego korzenia wrzątkiem (ok. 200-250 ml) i parz pod przykryciem przez 10-15 minut.
 Można pić 1-2 szklanki naparu dziennie.

id: wikibooks_31872
source: wikibooks
added: 2026-07-19
created: 2025-06-29
token_count: 543
license: CC-BY-SA-4.0
author: Wikibooks contributors, https://pl.wikibooks.org/wiki/Ksi%C4%85%C5%BCka_kucharska/Wieprzowina_curry

Składniki 
 40 dag schabu,
 1 średnia cebula,
 1½ łyżki przyprawy curry,
 szczypta imbiru w proszku,
 ½ l bulionu warzywnego,
 150 ml śmietanki kremówki 30% lub mleczka kokosowego,
 ½ łyżeczki słodkiej mielonej papryki lub papryki w płatkach,
 szczypta ostrej mielonej papryki,
 olej do smażenia,
 sól,
 pieprz,
 listki bazylii.

 Przygotowanie 
Mięso umyj, osusz i pokrój w niezbyt duże paski lub kostkę. Oprósz solą oraz pieprzem. Z papryki usuń gniazda nasienne, opłucz pod bieżącą wodą i osusz. Pokrój w słupki lub kostkę. Cebulę obierz z łupin, drobno posiekaj.
Na głębokiej patelni lub w rondlu rozgrzej olej, wrzuć mięso, posyp przyprawą curry i imbirem. Smaż, od czasu do czasu mieszając, aż lekko się zrumieni. Następnie dodaj cebulę, zeszklij. Wrzuć pokrojoną paprykę, całość smaż na […]

id: wikibooks_58505
source: wikibooks
added: 2026-07-19
created: 2023-12-10
token_count: 143
license: CC-BY-SA-4.0
author: Wikibooks contributors, https://pl.wikibooks.org/wiki/%C5%9Apiewnik/Umar%C5%82_Maciek%2C_umar%C5%82

Muzyka 
Autor: nieznany, melodia ludowa z Mazowsza

Źródło: 

 Tekst 
Autor: nieznany (piosenka ludowa)

Źródło: 

 Przypisy 

 Zobacz też 
 Idzie Maciek w serwisie Cyfrowa Biblioteka Polskiej Piosenki
 Idzie Maciek przez wieś w serwisie Antologia pieśni poskich Instytutu im. Oskara Kolberga
 Umarł Maciek, umarł w serwisie Stare Melodie (m. in. wykonanie Mieczysława Fogga)

🤖 Generated with Claude Code