Datasets:

SlayerLab
/

polish-dynaword

Adds sejm_committee_transcripts: 653,411 records and 166,310,754 measured cl100k_base proxy tokens (443,928,915 characters) from official Sejm committee full-zapis transcripts: 70 committees, 7,746 sittings, terms 9 and 10 (2019-11-14 to 2026-09-03). One record per speaker turn.

The source itself now ships in this PR: data/sejm_committee_transcripts/ (canonical parquet + per-record attribution sidecar + stats + datasheet + NOTICE + license-evidence snapshot), src/fetch_sejm_committee_transcripts.py (the full reproducible pipeline), src/test_sejm_committee_transcripts_contract.py, the sources.py entry with release: None and the artifacts/source_findings.md entry. Cross-check: rebuilding the shard from the pinned pipeline reproduces the token count exactly (166,310,754).

Register note: this is genuine multi-party discussion - committee examinations, ministry/expert Q&A, interjections ("Głos z sali"), not plenary addresses. Complements the existing plenary-only political/spoken sources.

What the texts are

Genre: official stenographic transcripts of standing-committee sittings - bill examinations, ministry and expert Q&A, chair announcements, votes, interjections. Register: prepared official transcript of spoken Polish - questions, answers and interruptions, formal address formulas, occasional read-out quotations of documents. Time span: sitting dates 2019-11-14 - 2026-09-03 (term 9: 331,411 records; term 10: 322,000), 11,261 distinct speaker labels. Not OCR'd (born-digital PDFs), not machine-translated.

Sample (verbatim, from the built shard)

Procedural turn by the chair:

id: sejm_committee_transcripts_9_SZA_77_119
source: sejm_committee_transcripts
added: 2026-09-18
created: 2022-01-26
token_count: 27
license: public-domain (official documents)
author: Przewodniczący poseł Marek Kuchciński (PiS)

Kto z państwa chciałby zabrać głos? Proszę bardzo. Pan poseł Nitras.

MP question and answer exchange (note the honest reconstruction limit: a mid-turn "Zastępca Szefa ..." label that the role-anchored pattern does not split):

id: sejm_committee_transcripts_9_RSP_60_36
source: sejm_committee_transcripts
added: 2026-09-18
created: 2022-10-26
token_count: 316
license: public-domain (official documents)
author: Poseł Jarosław Urbaniak (KO)

Raczej kwestia, czy nie ma kłopotów z zatrudnieniem osób z odpowiednimi kwalifikacjami i umiejętnościami przy obecnym poziomie wynagrodzeń? Takie problemy są zgłaszane w przypadku innych organów administracji centralnej.
Zastępca Szefa Kancelarii Senatu Karolina Zioło-Pużuk:
My też odnotowujemy to zjawisko i dotyczy ono wszystkich stanowisk, nie tylko stanowisk wysokospecjalistycznych, ale po prostu także stanowisk biurowych na poziomie
specjalisty czy głównego specjalisty. Zarobki oferowane przez Kancelarię, jako instytucję
publiczną, nie mogą jednak konkurować z zarobkami, które są na rynku, szczególnie
jeśli mówimy o Warszawie albo o kwestii dojazdów do Warszawy, bo przecież w stolicy
pracują osoby, które dojeżdżają nawet i 100 km. To jest faktycznie duży kłopot, ale niestety z żalem […]

Ministry answer:

id: sejm_committee_transcripts_9_RRW_42_127
source: sejm_committee_transcripts
added: 2026-09-18
created: 2021-04-13
token_count: 4706
license: public-domain (official documents)
author: Sekretarz stanu w MRiRW Anna Gembicka

Bardzo dziękuję. Szanowni państwo, dziękuję za wszystkie głosy i uwagi. Postaram się
odnieść do kwestii, które pojawiły się w trakcie dyskusji. A jeśli chodzi o kwestie, do których się nie odniosę, to oczywiście informację możemy przedstawić na piśmie.
Jeżeli chodzi o kwestie reformy zagospodarowania przestrzennego, to tak, jak pani
poseł też mówiła, to jest reforma zgłoszona przez inne ministerstwa. My oczywiście również jesteśmy zadowoleni z kierunku i z woli uregulowania tych kwestii. Sami widzimy
problemy, także w gminach wiejskich i miejsko-wiejskich.
Jeżeli chodzi o kwestie OZE, to inwestycje w OZE będą wspierane z funduszów spójności, a z kolei małe OZE – w ramach PROW. My też, jako ministerstwo, trochę przygotowując się do kolejnej perspektywy unijnej i do możliwości skorzystania […]

Before/after of the extraction and normalization

BEFORE is the raw pdftotext -layout output of a real zapis PDF (https://api.sejm.gov.pl/sejm/term10/committees/ASW/sittings/1/pdf), AFTER is the shipped record for the same turn (src/fetch_sejm_committee_transcripts.py: page-furniture removal, role-anchored turn reconstruction, soft-hyphenation joining, whitespace collapse, redact()):

BEFORE (raw layout text, first lines)

X kadencja

KANCELARIA SEJMU
Biuro Komisji Sejmowych

             PEŁNY ZAPIS PRZEBIEGU POSIEDZENIA
              ¢ Komisji Administracji i Spraw
                Wewnętrznych
                 (nr 1)
                 z dnia 21 listopada 2023 r.
[…]
W posiedzeniu udział wzięli pracownicy Kancelarii Sejmu: Anna Osińska – wicedyrektor Biura Komi-
sji Sejmowych, Magda Jedynak, Anna Ornat i Adrian Konefał – z sekretariatu Komisji w Biurze
Komisji Sejmowych.
Wicemarszałek Sejmu Krzysztof Bosak:
           Dobry wieczór.
              Przypadł mi zaszczyt poprowadzenia pierwszego posiedzenia Komisji Administracji
           Spraw Wewnętrznych. Witam wszystkich posłów. […]

AFTER (shipped record sejm_committee_transcripts_10_ASW_1_0, author: Wicemarszałek Sejmu Krzysztof Bosak)

Dobry wieczór.
Przypadł mi zaszczyt poprowadzenia pierwszego posiedzenia Komisji Administracji Spraw Wewnętrznych. Witam wszystkich posłów. Witam także przedstawicieli mediów. […]

The running headers ("KANCELARIA SEJMU" / "Biuro Komisji Sejmowych" / "PEŁNY ZAPIS ..."), page numbers and footers are dropped; hyphenation split across line breaks is joined (Komi-\nsji -> Komisji in the preamble example); turn text is kept verbatim otherwise. Full excerpt set in the review thread.

PII - which scrubber ran

Pattern pass redact() in src/fetch_sejm_committee_transcripts.py: e-mail -> [ADRES_EMAIL], checksum-valid PESEL/NIP/REGON and IBAN candidates -> [DANE_NUMERY], labelled phone/fax -> [DANE_KONTAKTOWE]. Counted in the shipped shard: 415 redactions (406 [ADRES_EMAIL], 6 [DANE_NUMERY], 3 [DANE_KONTAKTOWE]); the "454" in the original description was the pipeline's pre-dedup counter, which also counted rows later dropped as joint-sitting duplicates. This is not SlayerLab/NERGAL: the shard quotes citizens' statements and documents, so it must be re-run through NERGAL before release; merge with release: None is fine meanwhile. Stated in the datasheet.

Rights and method

Same LicenseRef-Polish-Official-Documents basis as sejm_api and sejm_interpellations: art. 4(2) of the Polish Copyright Act (official documents: https://eli.gov.pl/api/acts/DU/2025/24/text/O/D20250024.pdf) + Polish Open Data and Reuse of PSI Act, arts 2(12), 5, 6, 14, 15, 17 (https://eli.gov.pl/api/acts/DU/2023/1524/text.html). Provider: Kancelaria Sejmu RP via api.sejm.gov.pl (GET /sejm/term{n}/committees/{code}/sittings/{num}/pdf). No upstream Creative Commons grant claimed; evidence snapshot ships as sejm_committee_transcripts.license-evidence.json.

Method: pdftotext -layout; role-anchored speaker-turn reconstruction (validation: reproduces 98.9% of the earlier ASW pilot's audited corrected characters); body dehyphenation and page-furniture removal; sitting-level dedup of joint-committee sittings (74,117 rows / 912 duplicate sittings dropped); adjacent exact-turn dedup (42). Coverage: 8,658 of 9,175 term-9/10 sittings have a zapis PDF at source; 455 lack a PDF upstream, 62 failed after retries. Per-record provenance ships in sejm_committee_transcripts.attribution.jsonl: source_url pins the exact API PDF endpoint per row, content_sha1 fingerprints the transcript body.

Overlap and Slayer ontology

Committee sittings are a distinct register from plenary debates, so record-level overlap with parliamentary, ParlaMint-PL, sejm_api and sejm_interpellations should be near zero; residual risk: committee agendas and quoted speech can embed statute/print text, so near-dedup vs dziennik_ustaw/eurlex remains an integration gate. The Slayer ontology manifest carries content-addressed Version, Protocol with pinned script digests, Run, Evidence and 4 falsifiable Claims with Actors (in the source repo, linked below).

Benchmark overlap checks and controlled ablations remain pending. This PR proposes a source, not a merged or stable DynaWord release.

Test plan

python3 -m pytest src/test_sejm_committee_transcripts_contract.py — 12 passed
Rebuild cross-check: canonical rebuild reproduces 166,310,754 tokens / 443,928,915 chars exactly
Reviewer: texts are committee Q&A/interjections, not plenary speeches
Reviewer: statutory license basis accepted
Reviewer: NERGAL-before-release note accepted

PiotrSty changed pull request status to open 21 days ago

ppuzio

Fabryka AI org 11 days ago

Review against AGENTS.md ("What the PR must show", "Normalize the text before you build"). Needed before merge:

The PR has no commits. The data is only in the external PiotrSty/sejm-committee-transcripts repo. Push the source itself to this ref: data/<key>/ (parquet, stats, datasheet), fetch script, sources.py entry with release: None, contract test.
Sample inline. Paste 2–3 real documents into the description, one fenced block each: metadata as key: value lines, a blank line, then the text with its real line breaks, truncated with […] (format as in #24). A linked .sample.jsonl renders on the Hub as one escaped line per document and can't be reviewed at a glance.
PII. AGENTS.md requires SlayerLab/NERGAL. The source can merge with release: None after another scrubber, but it won't be admitted to a release until the shard is re-run through NERGAL. State in the datasheet which scrubber ran.

PiotrSty

Fabryka AI org 11 days ago

Review addressed — the source now lives in this PR: 63545ff on refs/pr/37 (single commit on main @ 34e7ea3). Point by point:

Commits pushed. data/sejm_committee_transcripts/ — canonical parquet (653,411 speaker turns; id, text, source, added, created, token_count, license, author), sejm_committee_transcripts.attribution.jsonl (per-record source_url pinning the exact API PDF endpoint + content_sha1, term/committee/sitting/turn_idx/speaker), .stats.json, datasheet, NOTICE.md, .license-evidence.json. Plus src/fetch_sejm_committee_transcripts.py (the full pipeline: API probe/download, pdftotext -layout parse, role-anchored turn reconstruction, dehyphenation, page-furniture removal, joint-sitting + exact-turn dedup, PII redaction, canonical build), src/test_sejm_committee_transcripts_contract.py, the sources.py entry with release: None, and the artifacts/source_findings.md entry. Two notes: the parquet is rebuilt into the canonical DynaWord schema (the external repo's train.parquet carried the pipeline's 10-column schema; per-record provenance moved to the attribution sidecar, which .gitattributes already LFS-tracks), and the rebuild cross-checks the numbers exactly — recomputed 166,310,754 tokens / 443,928,915 chars match the description above.
Sample inline. 3 real records pasted into the description, one fenced block each (key: value metadata, blank line, real line breaks, […] truncation, format as in #24): a chair procedural turn, an MP question + answer exchange (with an honest mid-turn label that the role-anchored split does not catch), and a ministry answer.
PII. Datasheet now states the scrubber: the pattern pass redact() in the committed script (e-mail -> [ADRES_EMAIL], checksum-valid PESEL/NIP/REGON + IBAN -> [DANE_NUMERY], labelled phone/fax -> [DANE_KONTAKTOWE]), counted in the shipped shard as 415 redactions (406 / 6 / 3). One correction to the description above: the "454" figure was the pipeline's pre-dedup counter, which also counted rows later dropped as joint-sitting duplicates. Agreed on the policy: not NERGAL, so no release before a NERGAL pass (this text quotes citizens' statements and documents); merge with release: None is fine meanwhile.

Also per AGENTS.md "What the PR must show" / "Normalize the text before you build": the datasheet describes the texts (official stenograms of spoken committee debate - genre, register, time span 2019-11-14 - 2026-09-03), the origin and the statutory license basis with both ELI links and a license-evidence snapshot; and before/after excerpts of the extraction are in the description (raw pdftotext -layout page with running headers and Komi-\nsji hyphenation vs the shipped turn). Every normalization step lives in the committed fetch script.

Verification: python3 -m pytest src/test_sejm_committee_transcripts_contract.py — 12 passed.

Add source: sejm_committee_transcripts (terms 9-10 committee zapis)667d9e7f

ppuzio changed pull request status to merged 10 days ago

ppuzio deleted the refs/pr/37 ref 10 days ago

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment