Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 87

What and why

Restores the licence version of every biblioteka_nauki row (slayerlabs/polish-dynaword-v2#57, v0.2.6 rights audit) and applies the maintainer's withhold decisions of 2026-10-06. The earlier recovery stored only the API's license.iconId (cc-by / cc-by-sa), which drops the version; v0.2.5 and the card said CC BY 4.0 / CC BY-SA 4.0 for all rows, while the API gives 2.0, 3.0, 3.0 PL and 4.0. The license column now holds the API's own licence per record. Nothing else in a row changes.

Baseline: Hub PR 62 (merged as 4e5009e; shard sha256 6b57b8eec198c9899198a73cea1af372df09b43783a49474a856ccbf83c97da4).

Maintainer decision, 2026-10-06 (applied in f7c7f32)

Approved by the maintainer in chat: "PR 81 withhold 12 404s + 56 NC/ND + 15 reprint rows, relabel articles_2011995 to CC BY-SA 4.0". The ids are the ones classified in follow-up sections 1, 3 and 4 below; anything not listed stays in.

Arithmetic: 42,055 (PR 81 at 2add5ca) + 1 (articles_2011995 back in) - 71 (56 + 15) = 41,985, which is what the build gives; 41,985 + 83 withheld = 42,068 (PR 62). The 71 ids were all among the 42,055 rows; none overlaps another list; none is unresolved in the sidecar. Tokens: 672,944,036 + 10,730 - 3,045,336 (2,715,677 + 329,659) = 669,909,430.

Mechanism: each list is an {id: reason} dict in src/clean_biblioteka_nauki.py, recorded in stats.json withheld like the API-side rows; the build exits with an error if a listed id (or the articles_2011995 override) is not among the rows it would otherwise keep, or if the API's licence for the override id gives another label. The contract test checks that none of the 71 ids is in the parquet, that articles_2011995 is labelled CC-BY-SA-4.0, that the reasons in stats.json equal the lists, and that the card counts per class equal stats.json. Not touched: README, CHANGELOG, make_docs.py (it skips custom datasheets and rewrites README/CHANGELOG), so the card's numbers were edited by hand from stats.json.

Kept by this decision (not listed by the maintainer): classes B, C and R3 to R5, and class (a) of section 1 (print colophon / boilerplate "all rights reserved", about 110 rows, except books_2028459). 4 of the 127 "all rights reserved" rows (whitespace-tolerant regex) are among the withheld rows; the shard now has about 120.

Rules applied

Withheld rows are listed with reasons in biblioteka_nauki.stats.json (withheld); none is kept under a guessed label.

New license distribution (generated from the shard, pinned by the contract test):

13,273 of the 41,985 rows are not 4.0. CC BY 28,280, CC BY-SA 13,705.

Rows and text unchanged (proof)

PR 62 shard vs this shard, run with a one-off comparison (prove_unchanged.py, not committed): 41,985 rows compared, same ids in the same order, SHA-256 of every text equal, every column except license equal. Digest over (id, sha256(text)) for both sides, PR 62 minus the 83 withheld ids: cc4ad9ce16aaef3c7bc5197743ad58182e3d5011f78cbf53896083e7a8541954. The digest function (content_digest) and a test that recomputes it from the shard are committed. Label transitions: cc-by -> 2.0: 1, 3.0: 2,413, 3.0-PL: 8,098, 4.0: 17,768, SA-4.0: 1 (articles_2011995); cc-by-sa -> 3.0: 7, 3.0-PL: 2,754, 4.0: 10,943. The shard was rebuilt twice from the pinned v0.2.5 file: same sha256 both times.

Provenance and licence summary

• Source of the licence: https://bibliotekanauki.pl/api/{articles|books|chapters}/{id}, field license (id, iconId, nameEn, urlEn, creativeCommons), one request per row, 2026-10-05/06, by data/biblioteka_nauki/recover_license_versions.py: one request clock (0.15 s), 3 workers, back-off on 429/5xx, stop on 403 or a non-JSON 200. 42,056 answered 200, 12 answered 404; no block, no CAPTCHA, no terms page. Responses are committed as biblioteka_nauki.license-api.json (sha256 51e0bf7900e016f4b0102eecfd5032f3969a566899bf8ba094abad9f4f4578b5), so the relabel reruns offline.
• Evidence file biblioteka_nauki.license-evidence.json (V/I marks, not_checked list, the maintainer decision); registry entry (traceable, provenance, legal_note) and card (new "Withheld rows" section, counts per class) updated. release stays "0.2.4". Provenance wording on the card, in the registry and in the contract test: "text from the SpeakLeash redistribution; licence and attribution recovered per document from the Biblioteka Nauki API (Hub PR 11); assembly script not in the repository" (commit 2add5ca, section 2). The Mixed-license limitation names the versions.
• Flags: the 5 shipped rows with reprintPublication (2) or translationPublication (3) are all still kept (none of the 71 is flagged); no row has restricted. Not used as a filter here.

Not decided, not verified

• Rows whose text contains "all rights reserved": the earlier review counted 128; not reproduced exactly (regex variants: 121, 126, 127, 137 on the 42,068 rows of PR 62). All of them have a CC BY / CC BY-SA licence in the API. Class (a) (print colophon / boilerplate) is kept; the footer-versus-API conflict for the rest is not resolved (follow-up, section 1: 25 notices read). 2 to 4 of the hits are among the withheld rows.
• Why the 12 records now answer 404, and whether the platform changed articles_2011995 after the first recovery: unknown.
• The v0.2.4 text-assembly script (build_parquet.py) cannot be found anywhere I looked; the card and registry now say so (follow-up, section 2). How the text layer was assembled is not re-verified.
• Attribution strings were not re-fetched.
• The 71 text-based ids come from regex hits plus one reader's pass; wordings the regexes do not match (other spellings, other languages) were not searched.

Follow-up measurements (made before the maintainer's decision; the decision is applied in f7c7f32)

1. "All rights reserved" notices: 25 of the 137-row superset read

Superset: the widest regex variant (NFKC text; wszelkie|wszystkie prawa zastrzeżone, all rights reserved, tolerant of line breaks and hyphenation) = 137 rows, 11,570,160 tokens (1.72%). Sample: random.Random(20261006).sample(sorted(ids), 25). For each row I read about 300 characters around every match, with the API licence beside it. One reader, one pass, not independently reviewed.

Classes: (a) the print edition's colophon or the publisher's (c) boilerplate; (b) a reservation written for this very text; (c) a reservation for embedded third-party material (figures, quoted works); (d) other.

• (a): 9 of the 20 are one journal (Topiarius, 107xxx), whose issue colophon says no part of the journal may be printed or reproduced without the publisher's consent. A string match finds that colophon in 46 of the 137 rows (proxy, not hand-read). Only 2 of the 20 print a CC statement in their own text: books_1182295 (CC BY-SA 3.0 PL, matches the API) and books_2028459 (CC BY-NC-SA 4.0, contradicts the API). For the other 18 the CC grant exists only as the API field. "Keep" rests on that field being the rights holder's declared licence on the platform; the colophon does read, literally, as a reservation on the whole publication. If the maintainer wants zero tolerance, drop all (a): about 110 rows, up to 1.7% of tokens.
• (b): articles_634916 is a translation of an Oxford University Press article, "przedrukowany za zgodą © Oxford University Press", with a printed licence number, "Artykuł z ograniczonymi prawami autorskimi. Zabrania się ... adaptacji, przeróbek i remiksów oraz publikowania, także w celach komercyjnych. Wszystkie prawa zastrzeżone". The API says CC BY-SA 4.0 and sets no translationPublication / reprintPublication flag. A phrase scan of the whole shard finds the same notice in exactly one other row, articles_634914 (same 2018 special issue of Wrocławski Rocznik Historii Mówionej; API CC-BY-SA-4.0). Two rows with this exact notice, 24,910 tokens; a wider scan (section 4) finds a third OUP row with different wording. The sample-based range for (b) is 1-27 rows; the phrase scan only finds this wording, so other (b) wordings could exist unseen.
• (d): articles_2109092 and articles_336963 are reference-list lines quoting a cited work's own (c) line; articles_483129 and books_2143725 are prose about "wszelkie prawa zastrzeżone" versus Creative Commons. None is a notice about these texts.
• Caveat on the intervals: rows come in series (9 of 25 are one journal), so independence is violated and the intervals are too narrow; no finite-population correction (25 of 137 sampled), which would narrow them by about 9%.

Finding outside the question: NC/ND licences printed in the text. Reading the notices turned up rows whose own front matter names a NonCommercial or NoDerivatives licence while the API says BY or BY-SA:

These 4 rows (892,611 tokens, 0.13%) are within the 137; the other 3 regex hits in the 137 are prose explaining licence types. Shard-wide, a narrow regex (a CC licence URL or the Polish licence name with NC or ND) matches 62 rows (3,031,883 tokens, 0.45%), 43 of them in the first or last 3% of the text. I read 12 of those 43: all 12 are the document's own licence statement (for example books_1396571 CC BY-NC 3.0 PL versus API CC-BY-3.0-PL; books_1409913 CC BY-ND versus CC-BY-4.0; books_1788278 CC BY-NC versus CC-BY-SA-3.0-PL; articles_1032116 "Attribution Noncommercial" versus CC-BY-2.0). The other 31 front/back-matter hits and 19 middle-of-text hits are unread. The other 58 hits are classified in section 3 (52 are the document's own NC/ND statement). The 4 rows above are withheld in f7c7f32.

2. build_parquet.py: not found; provenance wording fixed (commit 2add5ca)

Looked (read-only) and found nothing:

• git log --all -S"build_parquet": only mentions, no file. The original commit fbe1809 (Hub PR 11, 2026-08-03, author bartoszkobylinski1) ships recover_licenses.py, recover_attribution.py, the card, stats and Parquet; recover_licenses.py ends with "Next: build_parquet.py to assemble the CC-BY/SA subset".
• All 55 Hub refs of this dataset (4 branches, 6 tags, 45 PR refs): no build_parquet*, licenses.jsonl or index.jsonl.
• The contributor's Hub dataset repos (bartoszkobylinski1/aplikacje-prawnicze-mcq, pl-temporal): README only. SlayerLab/speakleash-raw-pl (4 files, README) and three other SlayerLab dataset repos: no biblioteka files.
• Hub PR 62 description and discussion (it says the script "was never committed"); Hub PR 11 description (it vendors only the two recover_* scripts); the card (links: bibliotekanauki.pl and the SpeakLeash key only); GitHub code search in the slayerlabs org and the v2 repo: no file.

Hub PR 11 says "The SpeakLeash blob has no per-document license", and that the parquet ships prebuilt because the live bibliotekanauki.pl probe returned HTTP 500 with zero records. The v0.2.5 card said "Pulled from SpeakLeash's public redistribution (speakleash-ds-pub, key biblioteka_nauki_pl_corpus)", and the registry's own file_key is that SpeakLeash key. The registry, however, already said (before this PR) "Fetched directly ... via the contributor rebuild script, not from the SpeakLeash redistribution", a specific claim that nothing in the repository or on the Hub supports (earlier I called it generic template wording; that was wrong). My first commit (7a28a14) aligned the card to the registry's claim, which made the card wrong too.

Fixed in commit 2add5ca (fast-forward 7a28a14..2add5ca, 4 files, no data change, shard sha256 unchanged): card Provenance section, registry entry and src/test_biblioteka_nauki_contract.py now say "text from the SpeakLeash redistribution; licence and attribution recovered per document from the Biblioteka Nauki API (Hub PR 11); assembly script not in the repository". The contract test fails if "direct rebuild" returns. Still not verified: how the text layer was assembled from the SpeakLeash blob. The PR 11 author is the only source for the script; I have not contacted anyone.

3. NC/ND statements in the text: all 62 regex hits classified

Input: bn_versions/nc_nd_text_hits.json (62 rows, narrow regex: a CC licence URL or the Polish licence name with NC or ND). The first 4 were classified in section 1; I then read the context of the other 58 (about 300 characters around each match, front or back matter of the document, with the API label beside it). One reader, one pass. At that point nothing was dropped or relabelled.

• Class A is 52 rows, 1,823,066 tokens (0.271% of the shard's 672,944,036). With the 4 rows from section 1 (books_2028459, books_2143725, articles_487021, books_1396563; 892,611 tokens) that is 56 rows, 2,715,677 tokens (0.404%) whose own text names an NC and/or ND licence while the API says BY or BY-SA. In every class-A row the printed statement contradicts the API label.
• Side finding, books_1788278 (216,594 tokens): the text is Nietyksza, "Ludność Warszawy", the attribution field is Demenko, "Klasyfikacja przebiegów częstotliwości podstawowej". The printed NC licence therefore belongs to a different work than the API record the label came from. Whatever the maintainer decides on its licence, its text/metadata alignment needs a look.
• Not resolved by reading: which statement the rights holder means. The API field is the platform's per-record declaration; the colophon is the print edition's. The more restrictive printed statement is the conservative reading; I applied it to nothing.
• Decision (maintainer, 2026-10-06): withhold class A and the 4 rows of section 1 (56 rows), keep B and C; applied in f7c7f32. Classes are one reader's judgement; the intervals of section 1 do not apply here because all 62 hits were read, but the regex itself only finds NC/ND written as a CC name or URL.

4. Publisher reprint-permission notices (shard-wide phrase scans)

Two scans over all 42,055 rows of the previous PR 81 head (NFKC text): (1) permission phrases in Polish and English next to "przedruk", "reprint", "translated", "za zgodą"; (2) any "by/with permission", "za zgodą/zezwoleniem" or "courtesy of" within 140 characters of a reprint, translation, publisher or (c) cue (131 rows, 8,382,306 tokens, nearly all bibliography entries or legal prose). I read the contexts of the 89 rows from scan 1 and of the edge-of-text hits (first or last 10%) from scan 2. Regex proxies, one reader: wordings not covered (other languages, "with the kind consent of") may exist.

• OUP-specific: 3 rows, 27,861 tokens. All three carry an API licence of CC-BY-SA-4.0. 27 more rows name OUP near a permission or licence word; the contexts I read are bibliography entries, a dictionary name and review headings, none a permission notice.
• Publisher or OUP permission on the whole text (R1 + R2): 15 rows, 329,659 tokens (0.049%). Counting R3 as well (author permission only): 20 rows, 400,417 tokens (520085 is also in A4 of section 3).
• None of the 35 rows in R1 to R5, and none of the 62 NC/ND rows, carries an API reprintPublication / translationPublication flag. The 5 flagged rows (books_1954543, books_2139603, articles_508678, articles_1194786, articles_2154637) were not found by these scans.
• Whether a CC BY-SA label is valid for a text printed with a publisher's reprint permission and "dalsze reprodukowanie wymaga zgody" depends on who granted the CC licence to the Polish translation; not decidable from the text. Decision (maintainer, 2026-10-06): withhold R1 and R2 (15 rows; articles_634916 also prints "Wszystkie prawa zastrzeżone"), keep R3 to R5; applied in f7c7f32.

pytest

Before (HEAD 2add5ca, script edited, data not yet rebuilt): 23 failed, 291 passed, 7 errors. After (f7c7f32): 23 failed, 294 passed, 7 errors (+3 tests: the maintainer-decision contract test, the override unit test, the withhold_listed unit test). The 30 failing or erroring test ids are identical before and after (none in biblioteka_nauki; they sit in other sources' tests, e.g. govpl, sejm_api_2011_2022, wiktionary_examples; cause not investigated here). src/test_biblioteka_nauki_contract.py: 39 passed.

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

id: biblioteka_nauki_articles_920161
source: biblioteka_nauki
added: 2026-06-27
created: 2010-01-01, 2010-12-31
token_count: 12049
license: CC-BY-4.0
attribution: Józef JAROSZ | 2010 | O ARCHAIZMACH JĘZYKOWYCH I PRAWNYCH ANACHRONIZMACH W KONSTYTUCJI KRÓLESTWA DANII

35 
 
 
O archaizmach językowych i prawnych anachronizmach w 
konstytucji Królestwa Danii 
 
Józef JAROSZ 
Uniwersytet Wrocławski 
Wydział Filologiczny 
Instytut Filologii Germaniskiej 
pl. Nankiera 15, 
50-140 Wrocław, Poland 
jaroj1@interia.pl 
 
Abstrakt: Niniejszy artykuł poddaje analizie tekst konstytucji Królestwa Danii, której treść 
i postać językowa bazują w dużej mierze na pierwszej duńskiej ustawie zasadniczej z 1849 roku. 
Badanie przebiega dwutorowo: obok uwag odnośnie języka ustawy ocenie poddano zawartość 
merytoryczną tekstu. Wnioski rozpatrywane są w odniesieniu do funkcji tekstu w ujęciu 
lingwistycznym oraz funkcji konstytucji w kontekście funkcji prawa. Badanie w sferze językowej 
wskazało na liczne archaizmy leksykalne, zawiłości składniowe i niedoskonałości […]

id: biblioteka_nauki_articles_151455
source: biblioteka_nauki
added: 2026-06-27
created: 2011-01-01, 2011-12-31
token_count: 7767
license: CC-BY-3.0
attribution: Z. Plutecki | 2011 | Wieloparametryczna ocena warunków komfortu cieplnego

PAK vol. 57, nr 2/2011    153 
 
Zbigniew PLUTECKI 
POLITECHNIKA OPOLSKA,  
ul. Sosnkowskiego 31, 45-272 Opole 
 
Wieloparametryczna ocena warunków komfortu cieplnego 
 
Dr inż. Zbigniew PLUTECKI 
 
Adiunkt na Wydziale Elektrotechniki, Automatyki  
i Informatyki Politechniki Opolskiej. Dorobek naukowy 
to 35 publikacji w czasopismach, krajowych i międzyna-
rodowych materiałach konferencyjnych, udział w 4 
projektach badawczych i 1 patent krajowy oraz ponad 
30 przemysłowych wdrożeń patentu. Członek Stowarzy-
szenia 
Polskich 
Energetyków. 
Główne 
kierunki 
działalności naukowej to optymalizacja układów 
energetycznych oraz badanie i modelowanie zjawisk 
cieplno-przepływowych. 
 
e-mail: z.plutecki@po.opole.pl 
 
 
Streszczenie 
 
W artykule przedstawiono […]

id: biblioteka_nauki_articles_555709
source: biblioteka_nauki
added: 2026-06-27
created: 2015-01-01, 2015-12-31
token_count: 4327
license: CC-BY-SA-4.0
attribution: Tomasz Rudowski | 2015 | SYTUACJA GOSPODARCZA W REGIONIE (STYCZEŃ – MARZEC)

GOSPODARKA 
 
 
Ameryka Łacińska, 1 (87) 2015, ss. 99-103 
99 
SYTUACJA GOSPODARCZA W REGIONIE 
(STYCZEŃ – MARZEC) 
REGION: WZROST GOSPODARCZY W 2015 ROKU NA POZIOMIE 2,2% 
AMERYKA ŁACIŃSKA I KARAIBY: STOPY REALNEGO WZROSTU PKB (2010-2015) 
Kraj/Region 
2011 
2012 
2013 
2014 
2015* 
Ameryka Łacińska 
i Karaiby 
4,5 
2,9 
2,7 
1,1 
2,2 
Ameryka Południowa 
Argentyna 
8,6 
0,9 
2,9 
-1,7 
-1,5 
Boliwia  
5,2 
5,2 
6,8 
5,2 
5,0 
Brazylia 
2,7 
1,0 
2,5 
0,3 
1,4 
Chile 
5,8 
5,5 
4,2 
2,0 
3,3 
Kolumbia 
6,6 
4,0 
4,7 
4,8 
4,5 
Ekwador 
7,8 
5,1 
4,5 
4,0 
4,0 
Paragwaj 
4,3 
-1,2 
13,6 
4,0 
4,5 
Peru 
6,5 
6,0 
5,8 
3,6 
5,1 
Urugwaj 
7,3 
3,7 
4,4 
2,8 
2,8 
Wenezuela  
4,2 
5,6 
1,3 
-3,0 
-1,0 
Meksyk i Ameryka Środkowa 
Kostaryka 
4,5 
5,1 
3,5 
3,6 
3,6 
Salwador 
2,2 
1,9 
1,7 […]

🤖 Generated with Claude Code