edukacja_medialna_pl: drop 4 lessons with embedded third-party quotations (252 -> 248)

#74
by ppuzio - opened
Fabryka AI org
•
edited 2 days ago

What and why

2026-10 rights audit for edukacja_medialna_pl (252 lessons, CC BY-SA 3.0, Fundacja Nowoczesna Polska): the audit found embedded third-party text "negligible" but listed "2-3 short quotes" and advised to "optionally remove the 3 quote passages". This PR drops the lessons that embed a quotation credited to a named third party, fail closed, in a committed script with a contract test. release stays None (admission waits for the NERGAL pass, slayerlabs/polish-dynaword-v2#60).

Before to after

before after
docs 252 248
tokens 728,987 713,052
chars 2,033,576 1,988,751
parquet sha256 dbce26650981e8e552c7fdac48bd35c02e9bd0a56fced4ea4871489b5a2365c4 1de35ab7c6c4db5fad67704c3f79afa65db6bfd411a1a55291bea05a4c940222

255 lessons were discovered; 3 were already rejected (no credited creator), now 7 (3 + 4 quote drops).

Rule: third_party_quote (4 lessons, whole lesson dropped)

src/clean_edukacja_medialna_pl.py (new). The raw XML is not in the checkout, so it filters the SHA-256-pinned PR #26 Parquet and sidecars (six pins, build() exits on any mismatch, on a DROP id missing from the input, on a marker that matches a lesson DROP does not list, and on a missing marker). The datasheet's "Reproduction" section gives the build, then this script, then a restore command for rerunning it from the PR #26 files. Dropped ids get selected: false, reason: third_party_quote in decisions.jsonl; attribution.jsonl, sample.jsonl (same SHA-ranked rule, 12), stats.json and qa.json are rewritten from the kept rows. Existing CRLF line endings are preserved, so the diff shows only the dropped records.

lesson id tokens status what it embeds
montaz-materialu-filmowego 2,942 confirmed a sentence quoted from dr hab. Wojciech Michera, credited to a University of Warsaw course ("cytat z materiałów do kursu")
scenografia-charakteryzacja-kostiumy-i-aktorzy 4,340 confirmed a second Michera sentence, same credit
kim-jestem-w-internecie 4,040 candidate invented-profile exercise with six "Ulubione cytaty" lines, each credited to a named person (Monroe, Churchill, ...)
wizerunek-w-sieci 4,613 candidate the same exercise, six quotations

Hand-check. The audit's own example (Michera in montaz-materialu-filmowego) is reproduced, plus the same quote type in scenografia-...: 1 passage each, 2 lessons. I could not reproduce a third. I read every passage in all 252 lessons that contains "cytat", a quotation followed by an attribution, or "za:" / "fragment ... tekstu", and found nothing else that embeds third-party text. The profile exercise is my guess at what the audit counted, dropped on the same fail-closed reasoning. The audit's count of 3 does not match what is in the data. Whether the two candidate lessons stay dropped is awaiting the maintainer's decision: see "Open decision: the two candidate lessons" below.

Kept on purpose: the attributed CC BY-SA Wikipedia excerpt in ekonomiczne-aspekty-wlasnosci-intelektualnej, the Constitution and ECHR article quotes (statutory text, Art. 4 PrAut), lessons that only name films or books, and uzaleznienia, which points to a handout by URL (the Jakubik article is not embedded).

Open decision: the two candidate lessons

Not decided yet. This PR currently drops kim-jestem-w-internecie (4,040 tokens) and wizerunek-w-sieci (4,613 tokens). The drop is fail-closed: they were my fail-closed guess, not a finding. Each carries six "Ulubione cytaty" lines in invented profiles, each credited to a named person. The audit says "2-3 short quotes"; I reproduced two (Michera) and still cannot show that these are the third, or that short attributed quotes inside an exercise need removing. Keeping them would be worth 2 documents and +8,653 tokens (250 documents, 721,705 tokens).

Recipe to keep both lessons:

git checkout 5ec3e18 -- 'data/edukacja_medialna_pl/*.parquet' 'data/edukacja_medialna_pl/*.jsonl' 'data/edukacja_medialna_pl/*.json'
# in src/clean_edukacja_medialna_pl.py: delete the two DROP entries that use PROFILES
python3 src/clean_edukacja_medialna_pl.py

Run in a scratch copy against the PR #26 inputs, the script with those two entries removed passes its pin and marker checks and gives 250 kept, 721,705 tokens, 2,013,259 characters, 5 rejected, and drops only the two Michera lessons. Not run, so not verified: the rest of the revert, which is manual. In src/test_edukacja_medialna_pl_contract.py set KEPT, TOKENS, CHARACTERS = 250, 721_705, 2_013_259, empty CANDIDATES and change the pinned 7 rejected to 5; then fix the counts and the four-lesson wording in the datasheet ("Third-party quotations", "Reproduction"), in the NOTICE.md sentence and in the artifacts/source_findings.md entry. The narrow git checkout pathspec above is also the one in the script's docstring and the datasheet; restoring the whole directory would bring back the old datasheet and NOTICE.md.

Provenance and licence

Unchanged by this PR. Lessons come from the Fundacja Nowoczesna Polska lesson XML and edukacjamedialna.edu.pl/lekcje/ pages; each kept lesson carries the per-lesson CC BY-SA 3.0 line and credited authors (62 credited authors across the 248 lessons). src/sources.py carries no counts, so the registry entry is untouched.

Other changes

  • data/edukacja_medialna_pl/edukacja_medialna_pl.md (counts, authors, "252 of 255" gate sentence, new "Third-party quotations" bullet naming the four slugs) and NOTICE.md (one sentence). A second commit (93efec9) extends the "Reproduction" section: after the build, run src/clean_edukacja_medialna_pl.py, which exits unless every input SHA-256 equals its pin in INPUT_SHA256 and the quotation markers match DROP; it also gives the restore command for a rerun from the PR #26 files. A third commit (2ee2c14) narrows the restore command in the clean script's docstring, which had restored the whole data directory (datasheet and NOTICE.md included), to the same three globs.
  • src/test_edukacja_medialna_pl_contract.py (new, 8 tests): the four ids absent from Parquet, attribution and sample; decisions selected: false with the reason; counts and tokens match stats; sidecars agree with the Parquet; sample rule reproduces the sample; no kept text contains a marker; build() on a miniature input drops the listed lesson, refuses a changed input, a stale id and an unlisted quotation, and refuses already-cleaned data; the "Reproduction" section names the clean script, INPUT_SHA256, DROP and the restore commit; the restore command is identical in the docstring and the datasheet and its globs match exactly the files in INPUT_SHA256.
  • src/test_edukacja_medialna_contribution.py (outside the usual touch list): it pinned the contributed data files through artifacts/edukacja_medialna_pl_ontology_manifest.json, which would fail on any data change. It now compares the manifest pins with the input pins in the clean script, so the manifest stays untouched. The pin of the build script is still checked. The manifest and build_edukacja_medialna_pl.py are unchanged: the manifest pins the builder by SHA-256, so the builder cannot take the clean step. The datasheet is not generated (five of its sections exist only in the .md, and the builder's README template would overwrite them), so its "Reproduction" section was edited directly.
  • artifacts/source_findings.md: appended entry.

Not established

  • The third quote the audit counts. The two "Ulubione cytaty" lessons are dropped fail-closed, pending the maintainer's decision; nobody has shown they must go.
  • That the builder reproduces the PR #26 files byte for byte. The raw XML is not in the checkout, so I could not run it. A rebuild that differs stops the clean script at its pin check instead of being filtered silently.
  • Whether the Michera sentences are short enough for the Art. 29 quotation right; dropping avoids the question.

pytest (worktree; LFS-dependent contract tests fail on pointer parquets)

  • before: 22 failed, 260 passed, 7 errors
  • after: 22 failed, 268 passed, 7 errors; the failure set is identical (sorted FAILED/ERROR lists diffed). test_edukacja_medialna_contribution.py passes before and after.

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

id: edukacja_medialna_pl_swiat-informacji-stoi-otworem
source: edukacja_medialna_pl
added: 2026-09-06
created: 2013-11-20
token_count: 3076
license: CC-BY-SA-3.0
author: Buchner, Anna; Dobrowolska, Urszula; Jasiewicz, Justyna; Joanna Góral-Wojtalik

Świat informacji stoi otworem!

Wiedza w pigułce

Dla nauczyciela

Osoba poszukująca informacji musi wybierać spośród wielu źródeł te, które są wiarygodne i użyteczne. Wyszukiwanie informacji należy rozpocząć od określenia potrzeby informacyjnej, tj. zastanowienia się, jakich informacji potrzeba, by zrealizować jakieś zadanie (np. przygotować referat lub dowiedzieć się, jak gdzieś dojechać).

Gdy uświadomimy sobie, czego chcemy się dowiedzieć, możemy rozpocząć poszukiwania informacji w różnych źródłach. Ze względu na stopień utrwalenia, źródła informacji możemy podzielić na dokumentalne (np. książki, prasa, filmy) i niedokumentalne (np. wypowiedzi, lekcje, przedstawienia).

Wyróżnić można również pośrednie źródła informacji --- a więc takie, które kierują użytkownika do źródeł na […]
id: edukacja_medialna_pl_wspolczesny-list
source: edukacja_medialna_pl
added: 2026-09-06
created: 2015-01-12
token_count: 3351
license: CC-BY-SA-3.0
author: Anna Walczak; Małgorzata Bazan; Urszula Dobrowolska; Wojciech Budzisz

Współczesny list

Wiedza w pigułce

W tym roku podczas wakacji byłem na koloniach plastycznych. Poznałem tam nowego kolegę, Bartka, którego bardzo polubiłem. Cały czas trzymaliśmy się razem. Niestety on był ze Szczecina, a ja pochodzę z Krakowa – nasze domy dzieli wiele setek kilometrów! Było mi bardzo smutno, kiedy obóz się kończył. Bartek jednak powiedział mi, że możemy łatwo utrzymywać kontakt przez internet.

– Ale w jaki sposób? – zapytałem kolegę.

– Sposobów jest wiele! – odpowiedział. – Możemy wysyłać sobie listy mailowe, rozmawiać na czacie, a także przez Skype'a. Wtedy nawet będziemy się widzieć!

– Ja nie mam jeszcze nawet swojej poczty e-mail...

– To nie szkodzi! Bardzo łatwo ją założyć i można to zrobić za darmo.

Bartek zapisał mi swój adres poczty e-mail na kartce. Przez […]
id: edukacja_medialna_pl_media-a-prawa-czlowieka-i-dziecka
source: edukacja_medialna_pl
added: 2026-09-06
created: 2012-11-09
token_count: 2962
license: CC-BY-SA-3.0
author: Baranowska, Joanna Ruta; Głowacka, Dorota; Paszewska, Weronika; Waglowski, Piotr

Media a prawa człowieka i dziecka

Wiedza w pigułce

Po II wojnie światowej powstała nowoczesna koncepcja praw człowieka. Zakłada ona, że podstawową, naturalną i niezbywalną wartością jest godność jednostki. Prawa człowieka regulują kontakt obywatela z państwem. Ich przestrzeganie zapobiega m.in. dyskryminacji. Na tym przekonaniu oparte są przepisy wielu krajowych i międzynarodowych aktów prawnych, które nazywają i porządkują prawa człowieka. Do najważniejszych należą:

Powszechna Deklaracja Praw Człowieka (1948)

Międzynarodowy Pakt Praw Obywatelskich i Politycznych (1966)

Europejska Konwencja Praw Człowieka (Konwencja o ochronie praw człowieka i podstawowych wolności 1953, w Polsce --- 1993)

Konwencja Praw Dziecka (1989)

Konstytucja Rzeczpospolitej Polskiej (1997)

Konwencja o […]

🤖 Generated with Claude Code

ppuzio changed pull request status to open
ppuzio changed pull request status to merged
ppuzio deleted the refs/pr/74 ref

Sign up or log in to comment