Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 85

What and why

Closes the global_voices item of slayerlabs/polish-dynaword-v2#50 (2026-10 rights audit).

Global Voices licenses "content created by Global Voices" under CC BY 3.0. Some posts are republications of a partner outlet's article, made under a content-sharing agreement, a partnership or with permission. Their text is the partner's work and the CC BY grant does not cover it. This PR drops those posts.

Commit d6bd578 on v026-global_voices, based on main c2099e6 (after #63). It changes these files:

• src/clean_global_voices.py, new. No build script existed, because the data were contributed directly in PR #7. The script filters the v0.2.5 parquet and attribution sidecar, each pinned by SHA-256, and rewrites stats.json. Kept rows are byte-identical and stay in their original order.
• src/test_global_voices_contract.py, new. It checks that no shipped row matches the rule and that the release has 1,926 rows. Parquet, stats and sidecar must agree on counts, ids, chars and tokens. Listed drops must be absent, and every hand-kept id must still match a rule, so stale exceptions fail.
• data/global_voices/: the parquet, attribution sidecar and stats.json are regenerated. The counts in the datasheet global_voices.md are updated, and it gets a new "Partner republications (excluded)" section. NOTICE.md drops Didem Tali's only recovered record (28444, an Ensia republication) and gains a note on the exclusion.
• src/sources.py: adds provenance and legal_note to the global_voices entry. release stays 0.2.3.
• artifacts/source_findings.md: one appended entry.

make_docs.py was not run. global_voices has custom_datasheet=True, so make_docs would not touch its datasheet anyway. It would also rewrite README.md, CHANGELOG.md and LICENSE, which this one-source PR must not edit. The datasheet counts were therefore edited by hand from stats.json.

The rule

The rule looks for the editorial republication note. It reads only the first 700 and the last 350 characters of a post, where that note sits. The text is split into sentences.

The self-referential rules (originally, republished, written_for, published_by, permission) fire only on a sentence that meets all three conditions:

• The sentence is about the post itself, for example "Ten artykuł…", "Jest to…" or "Poniżej publikujemy…".
• It does not name Global Voices or Rising Voices.
• Its subject is not an image.

permission is also skipped when the previous sentence is a photo caption, or when the consent comes from the author or artist. These conditions keep out news reporting such as "raport opublikowany przez UNESCO", GV's own projects and photo credits.

Each rule's count is the number of posts it matches; one post can match several rules.

Hand check

• All 114 drops were read. The English original was consulted where the Polish note was unclear. Per-document evidence is at the end of this description.
• 25 rule matches are kept with a stated reason.
  • 12 are cross-posts from the GV author's own blog: Juan Arellano's Globalizado ×7, Caribbean Free Radio, Hummus For Thought, Jillian C. York, Medium and r12n.
  • 7 were written by GV for the ISN blog.
  • 3210 is a GV post that another outlet republished.
  • 2 notes date GV's own English original (15986, 22228).
  • 3 are news mentions (8, 14503, 15837).
• 43 near-misses where the gating removed a raw regex hit were read; all are correctly kept. They are image credits, news reporting and GV projects.
• 15 partner-organisation accounts with no republication note were checked against the English page and kept. Examples are 350.org 22567 and Metamorphosis 33071. Nothing on the page marks them as republished.

Dropped posts by outlet (114 posts, 51 outlets)

Why 114 and not 64

The audit (artifacts/rights_audit_2026-10/reports/govweb.md) counted 64 posts that "say they were republished under a content-sharing agreement or partnership". It did not record its list or its script, so the two sets cannot be compared post by post.

• 67 of the 114 carry agreement or partnership wording. These are the posts matched by agreement or partnership, the class the audit counted. They include 20 PRI posts; the 21st PRI post, 28573, says only that it "początkowo pojawił się na łamach strony PRI.org" and has no agreement wording. The audit's "17 PRI" is therefore an undercount.
• 47 posts state the same fact without agreement wording. Their notes say "first published at X", "written for X", "published with permission" or "republished by non-commercial media". The audit's criterion would not catch them, but they are the partner's work just the same. By outlet:

The per-document table below gives each post's outlet, the rule that matched and the sentence it matched.

Borderline posts

These posts are dropped and flagged for review:

• Freemuse ×3. The note allows republication only by non-commercial media, so the licence is effectively NC, not CC BY.
• Bumaga 30003. A shorter version of the author's own article for Bumaga.
• Hive-Mind/EJO 34820.
• openDemocracy 29544. The note says "under a Creative Commons licence" without stating the version, which is likely NC. This was not verified for this article.
• Transitions 34277. The note says "powstał dzięki umowie partnerskiej", so the post may be a co-production rather than a republication.
• 32352. Built on two Meta.mk texts rather than a straight republication.

Two partner-organisation accounts with no note are kept: 350.org 22567 and Metamorphosis 33071.

Provenance and licence

The source is pl.globalvoices.org, contributed directly in PR #7 under CC BY 3.0. The Global Voices attribution policy scopes the licence to "content created by Global Voices", and its partnerships guide notes that GV "cannot republish their copyrighted content" for one-way republishing. Once the 114 partner posts are removed, the remaining posts are GV-authored or GV-translated. The datasheet and NOTICE.md already disclaim embedded photos and quotes. Per-document author and translator attribution stays in global_voices.attribution.jsonl, now 1,926 lines.

Input SHA-256:

• parquet 1ca02fdb…a45c3
• sidecar ab8ad393…fa13

Output SHA-256:

• parquet cd331337…c9c7ba
• sidecar a4709fcc…e691

Tests

python -m pytest -q, run from the worktree root:

• before, on main c2099e6: 23 failed, 240 passed, 7 errors
• after: 23 failed, 243 passed, 7 errors

The failure set is identical, verified with an empty diff of the FAILED/ERROR lines. These failures are pre-existing and come mostly from LFS pointers that are not pulled locally ("Parquet magic bytes not found"). The 3 new tests in src/test_global_voices_contract.py pass.

src/verify_paper_claims.py could not be run, and this is a pre-existing problem. Its default file.tex does not exist. Against the current paper main.tex, it raises LookupError: no table with \label{tab:s04-wersje} because the paper now uses tab:releases and tab:sources. It fails the same way at HEAD.

The paper's Global Voices rows (2,040 / 3,715,020 / 3.7M) describe v0.2.5. They need 1,926 / 3,453,918 / 3.5M after the v0.2.6 freeze. So do README.md line 208 and CHANGELOG.md line 170, which this one-source PR does not touch.

Sample

python3 src/render_sample.py data/global_voices/global_voices.parquet -n 3:

id: global_voices_1651
source: global_voices
added: 2026-07-16
created: 2010-02-28
token_count: 1078
license: CC-BY-3.0
author: Autorzy oryginału: Juliana Rincón Parra; Tłumaczenie: Krzysztof Pawliszak

Santiago po trzęsieniu (pviojo CC-By) Kiedy dzień chyli się ku końcowi, w Internecie pojawia się coraz większa ilość filmów ukazujących skutki dewastującego trzęsienia ziemi o sile 8.8 stopni w skali Richtera, które nawiedziło Chile o 3:30 rano (czasu lokalnego; 7:40 czasu polskiego). Trzęsienie ziemi, które objęło tereny na lądzie na skutek ruchów tektonicznych, dodatkowo wywołało fale tsunami. Na Pacyfiku ogłoszono alarm. Wiele krajów przygotowuje się na uderzenie fal. Samo trzęsienie ziemi zostało nagrane na wideo- przykładem jest ten materiał znaleziony na YouTube. Początkowe kadry pokazują pokój z komputerem, a kiedy ziemia zaczyna się poruszać, słychać dudnienie. Następnie gasną światła i zostajemy w ciemności, słysząc jedynie głos kobiety i mężczyzny, który ją uspokaja; […]

id: global_voices_28232
source: global_voices
added: 2026-07-16
created: 2015-10-07
token_count: 2520
license: CC-BY-3.0
author: Autorzy oryginału: Ndesanjo Macha; Tłumaczenie: Aleksandra Kling

Uchodźcy na węgierskiej autostradzie, w drodze na austriacką granicę. Zdjęcie: Joachim Seidler, licencja Creative Commons.
Prezydent RPA Jacob Zuma obwinił Zachód, w szczególności członków NATO, za ogromny napływ uchodźców do Europy.
W trakcie przemówienia, które wygłosił do zagranicznych dyplomatów w RPA we wtorek 15 września 2015 r., Zuma powiedział, że to akcja zbrojna przeciwko dyktatorowi Libii Muammarowi Kaddafiemu wywołała obecną falę uchodźców.
„To oni są za to odpowiedzialni. Oni wywołali tę sytuację i to oni muszą sobie z nią poradzić,” powiedział Zuma.
Przemówienie Zumy do dyplomatów można zobaczyć na tym filmie opublikowanym na YouTube przez ENCA News:
Zuma wezwał również do uznania Palestyny w granicach z 1967 roku, ostrzegając, że nie może być mowy o trwałym pokoju, dopóki […]

id: global_voices_20489
source: global_voices
added: 2026-07-16
created: 2013-10-31
token_count: 249
license: CC-BY-3.0
author: Autorzy oryginału: Rayna St.; Tłumaczenie: Alina Rychly

Crowdvoice jest platformą, która „umożliwia dostęp do sprawozdań z pierwszej ręki, czasami w połączeniu z najważniejszymi relacjami w mediach”. Poprzez uwydatnienie bitów dźwiękowych w celu zmiany świata na lepszy, Crowdvoice zestawia szeroki wachlarz źródeł w znaczącą całość, tym samym dając dostęp do złożoności historii, kryjących się za protestami. Najnowsza składanka przedstawia protesty w Iraku [ang.], obejmujące 10 lat zamieszek. Źródło jest niezwykle wartościowe dla każdego, kto próbuje dociec przyczyny ostatniej dramatycznej fali przemocy, która w samym tylko wrześniu 2013 spowodowała śmierć prawie 1000 osób [ang.]
Obraz osi czasu „Protesty w Iraku” na Crowdvoice.

🤖 Generated with Claude Code