Hugging Face's logo

Datasets:
SlayerLab
/
polish-dynaword

Tasks:
Text Generation
Languages:
Polish
Size:
1M<n<10M
ArXiv:
Tags:
polish
pretraining
dynaword
License:
Dataset card Files
xet
 Community
117
polish-dynaword / CHANGELOG.md
ppuzio's picture
ppuzio
Withdraw wikisource: one CC BY-SA label over mixed per-page licences (#60)
92b36dc
preview code
|
Raw History Blame
14.9 kB

Changelog

Unreleased (on main)

v0.2.6-dev plan

Sources planned for v0.2.6: on main, but no release admits them yet (release: None in src/sources.py). None has a recorded NERGAL pass, which is required before a release can admit it. Together: 15 sources, 1,285,412 documents / 1,256,790,015 cl100k-proxy tokens.

  • uzp_orzeczenia_pl by Piotr Styła: 34,715 documents / 422,043,881 tokens (Hub PR #39).
  • plwiki_user_talk by Piotr Styła: 380,527 documents / 381,445,111 tokens (Hub PR #40).
  • sejm_committee_transcripts by Piotr Styła: 653,411 documents / 166,310,754 tokens (Hub PR #37).
  • sejm_api_2011_2022 by Piotr Styła: 118,902 documents / 135,072,403 tokens (Hub PR #18).
  • plwiki_talk by Piotr Styła: 43,054 documents / 55,568,733 tokens (Hub PR #36).
  • ted_pl by Piotr Styła: 15,544 documents / 42,475,098 tokens (Hub PR #32).
  • sejm_interpellations by Paweł Puzio: 25,950 documents / 28,180,895 tokens (Hub PR #24).
  • nonsa by Piotr Styła: 12,304 documents / 16,883,883 tokens (Hub PR #41).
  • open_icm_pl by Piotr Styła: 85 documents / 2,358,174 tokens (Hub PR #34).
  • studia_bas by gzawadzki (PR #19, landed by Piotr Styła): 112 documents / 2,256,009 tokens (Hub PR #44).
  • wiadomosci_statystyczne_pl by Piotr Styła: 169 documents / 1,506,269 tokens (Hub PR #35).
  • rock_pollub_pl by Piotr Styła: 8 documents / 1,177,858 tokens (Hub PR #33).
  • edukacja_medialna_pl by Piotr Styła: 252 documents / 728,987 tokens (Hub PR #26).
  • open_agh_chemistry_pl by Piotr Styła: 370 documents / 564,028 tokens (Hub PR #25).
  • kultura_bez_barier_pl by Piotr Styła: 9 documents / 217,932 tokens (Hub PR #27).

Removed from main:

  • fandom_pl (Piotr Styła, Hub PR #48; 50,653 documents / 45,790,009 tokens from 12 wikis, never released and never counted in the totals above) was removed on 2026-10-05. Fandom's Terms of Use (revision 2025-12-19) bar scraping, bypassing robot exclusion and AI training without written consent; the pages were fetched live from the wikis' APIs on 2026-10-02. Samples also showed game text, lyrics and translated third-party stories under the CC BY-SA label. The registry entry, builder, test and ontology manifest were removed with the data.
  • european_hplt_v3_pl (1,603,974 documents / 2,555,525,652 tokens in v0.2.4 and v0.2.5) was removed by Arkadiusz Słota because its licence is unverified; it is re-added only after review. The registry keeps its release: "0.2.3" entry and make_docs.py skips it (no stats file). The v0.2.4 and v0.2.5 tags still contain it.
  • openstax_pl (Piotr Styła, Hub PR #23; 1,422 documents / 4,632,374 tokens, never released) was removed on 2026-10-05. OpenStax Poland's live prefaces allow only non-commercial reuse and bar LLM training without its written consent, and its CMS lists the eight books as CC BY-NC-SA 4.0; the fetcher had hard-coded CC BY 4.0. The registry entry, fetcher, test and ontology manifest were removed with the data.
  • common_voice_pl (Piotr Styła, Hub PR #22; 45,043 documents / 994,922 tokens, never released) was removed on 2026-10-05. CC0 is Common Voice's platform-level label: 24,709 of the sentences come from Europarl, whose European Parliament reuse notice requires source acknowledgement, which CC0 cannot carry, and the Sentence Collector sentences have no recorded sources. The registry entry, fetcher, contract test, ontology manifest and language audit were removed with the data.
  • wikiquote (30,363 documents / 31,896,591 tokens in v0.2.0-v0.2.5) was removed on 2026-10-05. The October rights audit labelled 158 of 200 randomly sampled pages as quoting works still in copyright; CC BY-SA covers the contributors' selection, not the quotes, which rest on the Art. 29 quotation right. As with HPLT, the registry keeps its release: "0.2.0" entry (with a legal_note) and make_docs.py skips it; plwikiquote was dropped from src/fetch_wikimedia.py. The v0.2.0-v0.2.5 tags still contain it.
  • wikisource (632,005 documents / 801,947,427 tokens in v0.2.0-v0.2.5) was removed on 2026-10-05. Every row carries one CC-BY-SA-3.0 label, but the October rights audit found in-copyright post-war imprints among 12 checked rows, each released by its rights holder under its own per-page licence (CC BY 3.0 PL, CC BY-SA 3.0 PL, CC BY-SA 3.0, GFDL, CC BY 4.0) and needing attribution the rows lack. The data came through SpeakLeash with an undocumented selection and no title, author or page id, so it cannot be screened. The registry keeps its release: "0.2.0" entry with a legal_note, and its traceable text no longer claims every text is public domain. A rebuild from the official dump with per-page licences is planned for a later release. The v0.2.0-v0.2.5 tags still contain it.

Changed on main:

  • samorzad_gov_pl rebuilt from the de-identified upstream v0.2 (Dawid Majewski, Hub PR #15, landed as #46): 72,771 documents / 41,154,506 tokens in v0.2.5 → 72,472 / 40,011,099 on main.
  • 1000_novels keeps only its public-domain texts, relabelled public-domain (Hub PRs #49, #55): 1,000 documents / 60,517,363 tokens in v0.2.5 → 837 / 51,765,528 on main.
  • wiktionary_examples dropped the lines quoting third-party works and is relabelled CC-BY-SA-4.0 (Hub PR #50): 8,639 documents / 1,115,820 tokens in v0.2.5 → 1,818 / 207,181 on main.
  • nkjp1m is relabelled CC-BY-4.0 per the NKJP Consortium, with an agreed ATTRIBUTION.md (Hub PR #51); documents and text unchanged.
  • The 17 v0.2.5 sources still on main total 2,093,026 documents / 6,276,340,429 tokens; all 32 sources on main total 3,378,438 documents / 7,533,130,444 tokens.

Tooling and process:

  • PII scrubbing moved out of the repo to SlayerLab/NERGAL (#38); contextual PII masking improvements (#28).
  • Releases are tagged (#31).
  • Source PRs are separate from releases; src/render_sample.py renders PR data samples (#42).
  • build_dynaword.py gained preserve_id / preserve_created flags (#45).
  • The v0.2.5 tag was re-pointed on 2026-09-29 from 02bcb0b to f71cc09, which corrects the v0.2.5 documentation to 20 sources. Every file under data/ is byte-identical between the two commits.

v0.2.5 (2026-08-14; tagged 2026-09-07)

  • The v0.2.5 tag contains 20 sources: samorzad_gov_pl, sejm_api and wiktionary_examples were added in this release. sejm_api and wiktionary_examples were previously documented as built but not released (release: None in src/sources.py); that was a documentation error, and the registry now sets release: "0.2.5" for both.
  • Added sejm_api by Piotr Styła: 38,812 parliamentary-speech documents (2023 onward) / 36,335,587 cl100k-proxy tokens from the official Sejm API, public-domain (official documents) (Hub PR #16).
  • Added wiktionary_examples by Katarzyna Majzel-Pośpiech: 8,639 Polish Wiktionary usage-example documents / 1,115,820 cl100k-proxy tokens, CC-BY-SA-3.0 (Hub PR #14).
  • Added the community contribution samorzad_gov_pl by Dawid Majewski (dawidmajewski): 72,771 Polish articles / 41,154,506 cl100k-proxy tokens from 246 local public institutions and one central platform tenant using the shared samorzad.gov.pl platform.
  • Converted the published, directly collected source dataset at revision d5eaefc32c3f17cfa5a01d14573f8e4ee7d43385; retained 72,771 of 81,422 records after dropping 6,624 short, 35 below the Polish-language heuristic and 1,992 exact-duplicate records.
  • Preserved the human-readable publishing institution in every author value. Added a 72,771-record attribution sidecar with discovered source URLs, fetched final URLs, recommended attribution URLs, titles, institution names, license URLs and modification notices.
  • Added the frozen 247-institution publisher manifest, reproducible conversion script, contract tests, source datasheet, NOTICE and release-time license evidence. The live shared footer script matched the audited SHA-256 b36ec507901ed443195479d8d8009386159ae59d7486ce266bfb3308043e99aa.
  • Recomputed complete-corpus totals: 4,366,651 documents and 9,676,513,980 cl100k-proxy tokens across 20 sources (earlier drafts of this entry gave 4,319,200 documents / 9,639,062,573 tokens across 18 sources, before sejm_api and wiktionary_examples were counted).

v0.2.4 (2026-08-07)

  • Expanded european_hplt_v3_pl (Arkadiusz Słota / Maggio33) from HPLT v3.0 WDS quality bins 10+9 to WDS bins 10-5: +1,493,384 documents / 2,043,141,874 cl100k-proxy tokens, 100% CC0-1.0. The source total is now 1,603,974 documents / 2,555,525,652 tokens.
  • Phone PII-scrub hardening (v12b): an independent-lens deep verification caught a systematic phone recall-gap (~1,900 real phone numbers) that the pipeline's self-report (idempotency) was structurally blind to — address-block suppression, dzwoń-label family, leading-0 landline, +48-paren, newline-over-span and foreign-UK-bare formats. Fixed and dual-independently verified; the residual 0.0135% is a classified clearance-tail. Phones map to a [Telefon] semantic tag (anti-poisoning, not a fake number).
  • Cross-source deduplication of the expansion against the existing WDS-10+9 shards: removed 3,038 near-duplicate documents (truncation variants); exact-dup 0, near-dup overlap 0.0000% (byte-verified). Resolves the "downstream near-dedup remains" caveat from the original release.
  • Contract-revalidated the merged parquet (canonical schema, non-empty text, positive token counts, uniform source/license, stats-file consistency).
  • Expanded contributor credit for Arkadiusz Słota (Maggio33) for the WDS 10-5 expansion, PII-scrub and dedup work.
  • Updated the complete-corpus release totals from per-source parquet stats: 4,246,429 documents and 9,597,908,067 cl100k-proxy tokens across 17 sources (also reflects the separately-merged biblioteka_nauki source now present on main).

v0.2.3 (2026-07-26)

  • Added the community contribution european_hplt_v3_pl by Arkadiusz Słota (Maggio33): HPLT v3.0 Polish WDS quality bins 10+9, 110,590 documents and 512,383,778 cl100k-proxy tokens under CC0-1.0 packaging.
  • Added the reviewed global_voices contribution by Dawid Majewski (dawidmajewski): 2,040 documents / 3,715,020 tokens with CC-BY-3.0 and per-document author attribution metadata.
  • Added nkjp1m from contributor 1am: 18,381 documents / 2,601,201 tokens from the manually annotated 1-million-word NKJP subcorpus under the upstream version-unspecified CC-BY statement.
  • Revalidated the contributed parquet against the canonical schema, non-empty text, positive token counts, uniform source/license metadata, and stats-file consistency.
  • Updated the complete 16-source release totals from per-source parquet stats: 2,710,974 documents and 6,881,277,362 cl100k-proxy tokens.
  • Recomputed the full-corpus phrase-frequency report and charts. Updated the Arrow reporting path to use 64-bit string offsets for large row groups.
  • Fixed src/test_hplt_v3_contract.py to resolve its fixture relative to the repository instead of a contributor-local Windows path.
  • Added public contributor credit for Arkadiusz Słota (Maggio33) and Paweł Puzio (ppuzio, author of the govpl contribution released in v0.2.2), plus Dawid Majewski (dawidmajewski) for Global Voices Polish.
  • HPLT caveat: CC0 applies to the HPLT packaging; underlying web documents can carry independent rights. Downstream near-dedup and web-content review remain recommended before training.

v0.2.2 (2026-07-19)

  • Added govpl — gov.pl Polish government press releases (CC-BY-SA-4.0), 88,190 documents / 80.7M tokens, fetched directly from gov.pl across 133 ministry/agency subsites. New scripts: src/discover_govpl.py (discovers each subsite's news-section path into the committed src/govpl_subsites.json manifest) and src/fetch_govpl.py (pure-stdlib crawl of that manifest).
  • Stable totals after adding govpl: 13 sources, 2,579,963 docs, 6.36B tokens (tiktoken proxy).
  • Article date range 2005-09-26 → 2026-07-19. Build dropped 24 non-Polish and deduped 1,926 cross-posted articles (exact sha1).

v0.3.1-preview (2026-07-10)

  • Added a separate 810-document Europeana direct-ingestion preview parquet at previews/v0.3.1/europeana.parquet. It was built from 1,000 live API records using per-record rights/creator metadata in 59.4 seconds; 808 retained records are Public Domain Mark and two are CC-BY-SA-4.0.
  • The direct source gate dropped 181 non-Polish and nine OCR-garble records. dsbench found unique IDs, no empty text, no duplicate text and no eval-set overlap. Potential-phone matches remain warnings for manual review.
  • This preview does not change v0.2.1 stable totals or promote Europeana to a stable source. It is the measured contract for the full rights-preserving run.

v0.3.0-preview (in progress)

  • Validated the source-policy gate on live local SpeakLeash shards: 12 reviewed sources sampled and cleared blocking checks; four sources remain excluded and two talk-page shards remain on hold pending a separate policy.
  • Ran a direct Europeana 200-record rebuild probe with per-record rights and creator metadata: 197 Public Domain Mark and three CC-BY-SA-4.0 records cleared blocking validation. This is candidate validation, not a release inclusion.
  • Biblioteka Nauki direct probe currently returns HTTP 500 and is held out of the preview until the upstream request is repaired and audited.

v0.2.1 (2026-07-02)

  • Released a stable metadata-schema update: id, text, source, added, created, token_count, license, author.
  • Current release totals after parquet recount: 12 sources, 2,491,773 docs, 6.28B tokens (tiktoken proxy).
  • Added 1000_novels as a stable CC-BY-4.0 literature source.
  • Added source registry entries and PR contract for future biblioteka_nauki and europeana ingestion with per-document license/author metadata. These sources are not included in v0.2.1 parquets yet.
  • Credited Bart Kobyliński for source expansion work on Biblioteka Nauki and Europeana metadata/rebuild preparation.
  • Kept raw SpeakLeash Europeana excluded unless a direct rebuild preserves per-record rights metadata.
  • Sources: eurlex, parliamentary, wikisource, wikipedia, dziennik_ustaw, wolne_lektury, 1000_novels, wikiquote, eltec_pol, wikivoyage, wikibooks, wikinews.
  • Excluded (see README): open_subtitles_corpus, europeana_eu_pl_corpus_raw_speakleash, project_gutenberg_pl_corpus.