README.md

metadata

license: cc-by-sa-4.0
language:
  - pl
pretty_name: Polish DynaWord
task_categories:
  - text-generation
size_categories:
  - 1M<n<10M
tags:
  - polish
  - pretraining
  - dynaword
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/*/*.parquet

Polish DynaWord

A continuously developed corpus of human-written Polish text, openly licensed or official, a Polish edition in the Dynaword family (Enevoldsen et al., arXiv:2508.02271).

v0.2.6 stable · 3,394,876 documents · 7.04B tokens (tiktoken proxy) · 27 sources Updated: 2026-10-22. v0.2.6 ships before the NERGAL personal-data pass; the pass runs on every source before v0.2.7 (Personal & sensitive data).

main, not a release (2026-10-22) · 27 sources · 3,394,876 documents · 7.04B tokens: the same data as v0.2.6. Three registered sources are held off main. Cite the v0.2.6 tag, not main.

v0.3-dev experimental track · quality/diversity workflow, source-gate validation and candidate-data audits. This is development work, not a released corpus version, and it does not replace the v0.2.6 stable parquets.

Europeana validation artifact (2026-07-10) · a separate reproducible 810-document direct-ingestion sample stored under the legacy path previews/v0.3.1/europeana.parquet. It is not a full dataset release.

Releases

Stable releases

version	status	documents	tokens	notes
v0.2.6	stable release	3,394,876	7.04B	Adds 12 sources and withdraws five; the 15 v0.2.5 sources it keeps are relabelled, filtered or rebuilt. Ships before the NERGAL personal-data pass.
v0.2.5	previous stable	4,366,651	9.68B	Adds community-contributed samorzad_gov_pl (institution-level author attribution and a URL-bearing attribution sidecar), sejm_api and wiktionary_examples.
v0.2.4	previous stable	4,246,429	9.60B	Expanded european_hplt_v3_pl to WDS bins 10-5 (+1.49M docs / 2.04B tok); phone PII-scrub hardening (independent dual-lens) + cross-source dedup (zero-overlap).
v0.2.3	previous stable	2,710,974	6.88B	Adds community-contributed european_hplt_v3_pl, global_voices and nkjp1m.
v0.2.2	previous stable	2,579,963	6.36B	Added community-contributed govpl: 88,190 docs / 80.7M tokens.
v0.2.1	previous stable	2,491,773	6.28B	12-source corpus with license and author metadata columns; added 1000_novels.
v0.2.0	previous stable	2,490,773	6.22B	Provenance-first corpus from 11 open/official sources.

The figures above are the releases as published. The tags v0.2.0-v0.2.5 were deleted after the October 2026 rights audit: they held data that v0.2.6 withdrew or changed (Withdrawn since v0.2.5, Changes since v0.2.5). Their notes stay below and in CHANGELOG.md; v0.2.6 is the earliest release that can be loaded.

Development and validation artifacts

name	type	scope	notes
v0.3-dev	experimental development track	workflow and candidate audits	Quality remix, legal-source caps, source QA, deduplication and direct-upstream ingestion. Not a corpus release.
Europeana validation artifact 2026-07-10	reproducible sample	810 documents / 140,042 tokens	Stored at the legacy path previews/v0.3.1/europeana.parquet; preserves per-record rights and creator metadata. Not a corpus release.

Version details

v0.2.6 — current stable

27 sources, 3,394,876 documents / 7,036,938,603 cl100k-proxy tokens:

• adds 12 sources, 1,097,821 documents / 637,130,408 tokens: plwiki_user_talk, sejm_committee_transcripts, plwiki_talk, ted_pl, sejm_api_2011_2022, nonsa, studia_bas, wiadomosci_statystyczne_pl, rock_pollub_pl, edukacja_medialna_pl, open_agh_chemistry_pl and open_icm_pl;
• keeps 15 of the 20 v0.2.5 sources, 2,297,055 documents / 6,399,808,195 tokens, each relabelled, filtered or rebuilt (Changes since v0.2.5);
• withdraws five v0.2.5 sources: european_hplt_v3_pl, wikisource, 1000_novels, samorzad_gov_pl and wikiquote (Withdrawn since v0.2.5);
• leaves out sejm_interpellations and kultura_bez_barier_pl, both held off main.

v0.2.6 ships before the NERGAL personal-data pass. Only each source's own filters apply (see its datasheet); the pass runs on every source before v0.2.7.

v0.2.5 — previous stable

Added three sources, bringing the complete release to 20 sources, 4,366,651 documents / 9,676,513,980 cl100k-proxy tokens:

• samorzad_gov_pl: 72,771 documents / 41,154,506 cl100k-proxy tokens from 246 Polish local public institutions and one central platform tenant. The contribution preserves the publishing institution as author, discovered and fetched article URLs in an attribution sidecar and the verified platform-wide CC-BY-SA-4.0 basis.
• sejm_api: 38,812 documents / 36,335,587 cl100k-proxy tokens of parliamentary speeches (2023 onward) from the official Sejm API, public-domain (official documents) (PR #16).
• wiktionary_examples: 8,639 documents / 1,115,820 cl100k-proxy tokens of Polish Wiktionary usage examples, one document per headword, CC-BY-SA-3.0 (PR #14).

Documentation correction: the v0.2.5 release has all 20 sources. Earlier copies of this card counted only 18 (4,319,200 documents / 9,639,062,573 tokens) and listed sejm_api and wiktionary_examples as not released.

v0.2.4 — previous stable

Expanded european_hplt_v3_pl from WDS bins 10+9 to WDS bins 10-5 (added quality bins 8/7/6/5 and further WDS-9 shards): +1,493,384 documents / 2,043,141,874 cl100k-proxy tokens, 100% CC0-1.0. Includes a phone PII-scrub hardening pass (v12b) verified by two independent lenses (a systematic recall-gap masked by pipeline self-report was caught pre-release), and cross-source deduplication against the existing WDS-10+9 shards (exact-dup 0, near-dup overlap 0.0000%, byte-verified) — resolving the "downstream near-dedup remains" caveat from the original release.

v0.2.3 — previous stable

Community expansion release. Adds european_hplt_v3_pl from PR #5 and global_voices from PR #7, plus nkjp1m from PR #8. The complete 16-source release contains 2,710,974 documents / 6,881,277,362 cl100k-proxy tokens, summed from the released per-source parquet statistics.

v0.2.2 — previous stable

Added govpl from PR #9: 88,190 Polish government press releases collected directly from 133 gov.pl ministry and agency subsites.

v0.2.1 — previous stable

Introduced the canonical eight-column release schema: id, text, source, added, created, token_count, license, author. Added 1000_novels from PR #1 and recomputed all release statistics from parquet files.

v0.2.0 — initial stable corpus

The provenance-first baseline: 11 reviewed open or official sources, 2,490,773 documents and 6.22B proxy tokens.

v0.3-dev — experimental quality workflow

Candidate-only work on a better-balanced training mixture: legal-source caps, temperature sampling, source QA, deduplication and direct-upstream ingestion. It is not a replacement for the stable corpus.

Europeana validation artifact — 2026-07-10

A separate reproducible 810-document Europeana sample preserving per-record rights and creator metadata. It does not change stable-release totals. v0.3.1-preview remains only as the legacy storage-path label.

What this dataset contributes

The raw texts come from upstream publishers, official sources and existing open corpora; eurlex, wolne_lektury and biblioteka_nauki come from SpeakLeash's redistribution. The value added here is the curation, not the bytes, following the Dynaword methodology:

1. License review per source — each source vetted for a traceable legal basis (a licence, the public domain or official-document status, documented in its datasheet); sources that fail the review are excluded with a stated reason (see table below), not silently kept. This is the core editorial work.
2. Filtering & normalization — minimal, reproducible gates (short-doc, non-Polish, exact cross-source dedup, OCR garble) applied uniformly to one clean schema: id, text, source, added, created, token_count, license, author.
3. Documentation — a datasheet per source (Gebru et al. 2021) + this card, so provenance and licensing are auditable rather than assumed.
4. Reproducibility & versioning — src/ rebuilds the corpus from sources; new sources and removals are tracked in the CHANGELOG.

Credit for the underlying texts belongs to their authors and upstream sources, and to SpeakLeash as the redistributor of eurlex, wolne_lektury and biblioteka_nauki; this release does not claim ownership of them (see Disclaimer).

Contributors

contributor	contribution	release / PR
Kacper Wikieł	Project maintainer; corpus curation, source and license review, release engineering, documentation, validation and reproducible build workflow.	all releases
Bart Kobyliński	Added 1000_novels; expanded Biblioteka Nauki and Europeana ingestion with per-document license and author metadata.	v0.2.1, PR #1
Paweł Puzio	Built govpl: subsite discovery, direct ingestion pipeline, dataset artifact, contract tests and documentation. Added sejm_interpellations (held off main, not yet released); contributor guidance, PII scrubbing and its move to SlayerLab/NERGAL; registry-driven release totals, release tags and the source-PR/release split.	v0.2.2, PR #9; #17, #20, #24, #28–#31, #38, #42
Arkadiusz Słota	Built & expanded european_hplt_v3_pl: HPLT v3 WDS 10+9 (PR #5) then WDS 10-5 expansion (+1,493,384 docs / 2.04B tok), modular cleaning pipeline, phone-recall PII-scrub (v12b, independent dual-lens), cross-source dedup-vs-base (zero-overlap), validation and documentation.	v0.2.3 (PR #5), v0.2.4
Dawid Majewski	Added the reviewed Global Voices Polish corpus; built samorzad_gov_pl with direct acquisition, institution attribution, licensing evidence, release artifact and reproducible conversion; prepared its de-identified upstream v0.2.	v0.2.3, PR #7; v0.2.5, PR #13; PR #15 (merged via #46)
1am	Added nkjp1m: the manually annotated 1-million-word NKJP subcorpus, direct fetch/build pipeline, source documentation and release artifact.	v0.2.3, PR #8
Piotr Styła	Corpus preparation and contribution of sejm_api: parliamentary speeches from the official Sejm API with speaker attribution and immutable PiotrSty/sejm-speeches-corpus v1.0.1 provenance. Added 11 sources released in v0.2.6 (sejm_api_2011_2022, open_agh_chemistry_pl, edukacja_medialna_pl, ted_pl, rock_pollub_pl, open_icm_pl, wiadomosci_statystyczne_pl, plwiki_talk, sejm_committee_transcripts, plwiki_user_talk, nonsa) and kultura_bez_barier_pl (held off main, not yet released); uzp_orzeczenia_pl, held off main until v0.2.8; landed studia_bas and the samorzad_gov_pl v0.2 update; build_dynaword preserve_id/preserve_created flags.	v0.2.5, PR #16; v0.2.6, #18, #25–#27, #32–#37, #39–#41, #44–#46
Katarzyna Majzel-Pośpiech	Added wiktionary_examples: usage examples extracted from the Polish Wiktionary dump, one document per headword.	v0.2.5, PR #14
gzawadzki	Added studia_bas: peer-reviewed socioeconomic and public-policy articles from Studia BAS.	v0.2.6, PR #19 (merged via #44)

Contributions are credited when they add a verifiable dataset artifact, pipeline, validation, documentation or release work. A merged discussion with no resulting files is not listed as a data contribution.

Guiding principles

1. Open & traceable licensing — every source is openly licensed or outside copyright, with a documented legal basis (see each datasheet's "traceable basis"), not a vague "public domain". A few sources have a contested legal status rather than a licence; their datasheets describe it.
2. Reproducibility — src/build_dynaword.py rebuilds the corpus from sources.
3. Documented — a datasheet per source under data/<source>/.
4. Extensibility — versioned; new sources via PR.

Sources

source	description	license	documents	tokens
eurlex	EUR-Lex (EU legal acts, Polish)	LicenseRef-EU-Reuse-Policy	242,488	2,333.0M
parliamentary	Polish Parliamentary Corpus (Sejm/Senat)	public-domain (official documents)	329,863	1,504.1M
wikipedia	Polish Wikipedia	CC BY-SA (3.0, 4.0 or both per record, by contribution date)	1,468,647	1,226.9M
biblioteka_nauki	Biblioteka Nauki	per-record upstream license	37,843	590.9M
dziennik_ustaw	Dziennik Ustaw (Polish primary legislation)	public-domain (official documents)	35,442	486.1M
plwiki_user_talk	Polish Wikipedia user-talk pages (ns 3)	CC BY-SA (3.0, 4.0 or both per record, by contribution date)	380,295	381.0M
sejm_committee_transcripts	Sejm committee transcripts (terms 9-10)	public-domain (official documents)	613,722	112.0M
wolne_lektury	Wolne Lektury (school readings)	per-record Wolne Lektury licence (public domain, LAL 1.3, CC BY-SA 3.0 PL/3.0/4.0)	5,900	91.7M
govpl	gov.pl — Polish government press releases	CC-BY-SA-4.0	74,946	68.3M
plwiki_talk	Polish Wikipedia talk pages (ns 1+5)	CC BY-SA (3.0, 4.0 or both per record, by contribution date)	43,054	55.6M
ted_pl	TED Polish public-procurement notices	TED notice status, no licence granted	14,210	39.6M
sejm_api	Sejm API parliamentary speeches (2023 onward)	public-domain (official documents)	39,776	30.6M
sejm_api_2011_2022	Sejm speeches 2011-2022 (terms VII-IX, attributed)	public-domain (official documents)	33,930	26.9M
eltec_pol	ELTeC-pol (European Literary Text Collection, Polish)	CC-BY-4.0	100	21.5M
nonsa	Nonsensopedia (Polish satirical wiki, content namespaces)	CC BY-SA 3.0	11,692	15.4M
wikibooks	Polish Wikibooks (open textbooks)	CC BY-SA (3.0, 4.0 or both per record, by contribution date, plus the licence of imported text)	9,041	15.2M
wikivoyage	Polish Wikivoyage (travel guides)	CC BY-SA (3.0, 4.0 or both per record, by contribution date)	7,312	14.1M
wikinews	Polish Wikinews	CC BY (2.5, 4.0 or both per record, by contribution date, plus the licence of imported text)	23,572	11.0M
global_voices	Global Voices Polish	CC-BY-3.0	1,926	3.5M
nkjp1m	The manually annotated 1-million word subcorpus of the National Corpus of Polish	CC-BY-4.0	18,381	2.6M
studia_bas	Studia BAS — peer-reviewed socioeconomic and public-policy articles	CC-BY-3.0-PL	112	2.3M
wiadomosci_statystyczne_pl	Wiadomosci Statystyczne. The Polish Statistician (GUS)	CC BY-SA 4.0	169	1.5M
rock_pollub_pl	ROCK Politechnika Lubelska - Polish academic books	CC-BY-SA-4.0	8	1.2M
edukacja_medialna_pl	Edukacja Medialna - Polish media-literacy lessons	CC-BY-SA-3.0	248	0.7M
open_agh_chemistry_pl	Open AGH - four Polish chemistry textbooks	CC-BY-SA-4.0	370	0.6M
open_icm_pl	OPEN ICM Polish open-access publications pilot	per-record CC BY or CC BY-SA	11	0.4M
wiktionary_examples	Polish Wiktionary usage examples	CC BY-SA (3.0, 4.0 or both per record, by contribution date)	1,818	0.2M
total			3,394,876	7,036.9M

Each document's license field gives its own licence or status; the label above summarises the source.

Withdrawn since v0.2.5

These v0.2.5 sources are not in v0.2.6, and no release tag holds them: the earlier tags were deleted (Releases).

source	v0.2.5 documents	v0.2.5 tokens	reason
european_hplt_v3_pl	1,603,974	2,555.5M	Licence unverified: HPLT's CC0 covers its packaging, not the crawled pages. It will not be re-added.
wikisource	632,005	801.9M	Its single CC BY-SA 3.0 label also covered in-copyright post-war imprints that carry their own per-page licences. A rebuild from the official dump is planned for a later release.
1000_novels	1,000	60.5M	Its rights screen (PR #49, PR #55) kept the public-domain texts, but no row records the edition its text follows, so a protected preface, abridgement or editor's corrections cannot be ruled out. Withdrawn on 2026-10-17 (PR #124).
samorzad_gov_pl	72,771	41.2M	Its CC BY-SA 4.0 is the platform operator's footer, not a grant from the 247 institutions that wrote the texts. Withdrawn on 2026-10-17 (PR #123).
wikiquote	30,363	31.9M	Most of its pages quote works still in copyright, and its CC BY-SA label covers only the contributors' selection.

Changes since v0.2.5

The 15 v0.2.5 sources in v0.2.6 total 2,297,055 documents / 6,399.8M tokens.

• eurlex is relabelled LicenseRef-EU-Reuse-Policy: v0.2.5 labelled it CC-BY-4.0, but the EUR-Lex notice grants CC BY 4.0 only for consolidated texts and editorial content, not the acts themselves, and excludes third-party material. The 92 documents carrying the IASB notice and 480 more found by a row review (third-party credits, UN ECE regulations, European Parliament stubs) are dropped, and eurlex.attribution.jsonl gives each row its EUR-Lex link. 242,488 documents / 2,333.0M tokens (PR #71, PR #82, PR #173, PR #177).
• parliamentary was rebuilt from IPI PAN's speaker-labelled TEI dump: only official speakers are kept in the sittings and exact copies are dropped. The System Informacyjny Sejmu credit is given for its Sejm material as a precaution. 329,863 documents / 1,504.1M tokens (PR #108, PR #111).
• wikipedia was rebuilt from the official plwiki dump, replacing the SpeakLeash snapshot. Each record is labelled CC BY-SA 3.0, 4.0 or both by its contribution dates. The 35 pages that carry plwiki's copyright-violation notice are dropped, and so are 7 pages that plwiki deleted for a copyright violation. 1,171,897 → 1,468,647 documents, 707.2M → 1,226.9M tokens (PR #68, PR #128, PR #148).
• biblioteka_nauki keeps its text from the SpeakLeash redistribution. Each record now carries the licence version and licence basis the Biblioteka Nauki API gives for it. 3 records under the Public Domain Mark and 4,142 further rows were dropped after review. 42,071 → 37,843 documents, 673.5M → 590.9M tokens (PR #62, PR #81, PR #186).
• dziennik_ustaw no longer claims Monitor Polski, which it never included. A sidecar gives each row its ELI address, and the System Informacyjny Sejmu credit is given as a precaution. Its text and counts are unchanged: 35,442 documents / 486.1M tokens (PR #103, PR #160, PR #162).
• wolne_lektury is relabelled per document: v0.2.5 labelled every document CC-BY-SA-4.0 / Wolna Sztuka 1.3, which is wrong. Each now carries the licence its own Wolne Lektury book states (4,383 public domain, 1,467 LAL 1.3, 50 CC BY-SA 3.0 PL, 3.0 or 4.0), with its credits in wolne_lektury.attribution.jsonl. 67 documents are dropped (63 eltec_pol novels, 2 with no usable licence, 2 whose translation is not free), 174 printed in two later editors' selections are held out until the books' contents are checked, and the author's and translator's notes are added at the end of 459 texts. 5,900 documents / 91.7M tokens (PR #67, PR #99, PR #184, PR #188, PR #190, PR #191).
• govpl drops 188 PAP reprints, 186 other third-party texts and 12,870 press releases of prosecutor's offices. 88,190 → 74,946 documents, 80.7M → 68.3M tokens (PR #63, PR #170, PR #214).
• sejm_api cuts the "Przebieg posiedzenia" trailers, puts the 2,466 statements they appended into rows of their own and drops 123 rows by speakers outside its allowlist. 38,812 → 39,776 documents, 36.3M → 30.6M tokens (PR #83, PR #158).
• eltec_pol adds author metadata and an attribution sidecar and restores the chapter titles: 100 documents / 21.5M tokens (PR #187).
• wikibooks withholds 54 pages: 51 by the rights review and 3 with {{GFDL-tekst}}. Each record is labelled by its contribution dates and the licence of imported text, and markup residue is removed. 9,112 → 9,041 documents, 15.6M → 15.2M tokens (PR #127, PR #134, PR #139, PR #142, PR #147).
• wikivoyage withholds 6,330 pages: 6,326 with {{Atrybucja}} (text from Wikitravel), 2 with {{Import}} and 2 by the rights review. Each record is labelled by its contribution dates, and markup residue is removed. 13,645 → 7,312 documents, 17.1M → 14.1M tokens (PR #125, PR #133, PR #138, PR #141).
• wikinews withholds 1 page. Each record is labelled by its contribution dates and the licence of imported text, and markup residue is removed. 24,386 → 23,572 documents, 12.1M → 11.0M tokens (PR #77, PR #130, PR #132, PR #137, PR #146).
• global_voices drops 114 posts republished from partner outlets. 2,040 → 1,926 documents, 3.7M → 3.5M tokens (PR #66).
• nkjp1m is relabelled CC-BY-4.0 per the NKJP Consortium and ships an agreed attribution file; its text is unchanged (PR #51).
• wiktionary_examples dropped the lines quoting third-party works and is relabelled per record CC-BY-SA-3.0, CC-BY-SA-4.0 or both, by contribution date. 8,639 → 1,818 documents, 1.1M → 0.2M tokens (PR #50).

Held off main

These registered sources (release: None) have no data on main and are in no release (see CHANGELOG.md):

• sejm_interpellations, until v0.2.7: it quotes letters from the public, some with health details, and waits for the NERGAL pass.
• kultura_bez_barier_pl, until the scope of its licence is answered.
• uzp_orzeczenia_pl, until v0.2.8: the personal-data pass leaves part of the sole traders' firm names and addresses unmasked.

Method

Only human-authored text — no synthetic, machine-translated, or auto-transcribed data. Gates are intentionally minimal (drop short docs, non-Polish, exact duplicates, OCR garble); heavy quality filtering and mix-weighting are left to downstream training. Evaluation-set decontamination is applied/marked separately. Schema: id, text, source, added, created, token_count, license, author. The license and author columns are per-document metadata when upstream exposes them; older sources use the source-level license and an empty author field.

Cross-source deduplication covers some pairs of sources only. Wolne Lektury drops a document when at least 0.9 of its 5-word shingles are in one ELTeC novel; below that, both copies stay. Both tomes of Helena Mniszek's Trędowata (0.7588 and 0.6473 against the same, shorter ELTeC text) are therefore in both wolne_lektury and eltec_pol; the pairs are listed in data/wolne_lektury/wolne_lektury.eltec_pairs.tsv.

v0.3 quality roadmap and current status

The v0.2.x raw corpus is intentionally provenance-first, but its token mix is too heavy in legal/parliamentary language for natural general pretraining. The v0.3 workflow therefore separates source inclusion from training mix:

• cap eurlex + parliamentary + dziennik_ustaw to roughly 10-20% of training tokens combined;
• use source-level temperature sampling (sqrt, alpha 0.5) instead of raw token-proportional sampling;
• add traceably licensed contemporary/natural Polish: open web, academic prose, cultural heritage text, guides, technical documentation/blogs, Q&A, and dialogue/instruction data;
• run aggressive exact, normalized, and near-duplicate removal;
• reserve the final 5-15% of training for higher-quality sources rather than the largest sources;
• evaluate per-source perplexity and style contamination, not only global loss.

Current v0.3 source-ingestion status:

• biblioteka_nauki: included with its text from SpeakLeash's redistribution and each document's licence and author from the Biblioteka Nauki API.
• europeana: prepared in the source registry as a direct-upstream rebuild target with per-record rights statements and creator metadata; raw SpeakLeash Europeana remains excluded.
• Europeana release policy: split conservatively at pre-1929 records for US-sensitive downstream reuse, and keep later/unknown records separately labeled or held until their US rights status is checked.
• european_hplt_v3_pl: WDS bins 10-5 shipped in v0.2.4 and v0.2.5, then the data was removed from main because its licence is unverified: HPLT packaging is CC0, but the underlying crawled web documents can carry independent rights. It will not be re-added.
• 1000_novels: a per-text rights screen kept the texts in the public domain in Poland, but no row records the edition its text follows, so the source was withdrawn from main on 2026-10-17. Re-admission, per text with the edition identified, is tracked for v0.2.8.

Current review artifacts:

• configs/source_candidates_v0_3.json — candidate decisions and license policy.
• artifacts/source_license_review_v0_3.md — source-by-source license review.
• artifacts/source_candidate_audit_v0_3.md — generated Hugging Face metadata audit.
• artifacts/training_mix_v0_3.md — example 1B-token training mix with legal sources capped at 15%.
• artifacts/bartek_source_ingestion_plan_2026-07-02.md — PR contract for Biblioteka Nauki and Europeana ingestion.

Excluded sources (transparency)

Sources we reviewed and deliberately left out — part of the curation:

source	reason
open_subtitles_corpus	Derivative of copyrighted film/TV dialogue; OpenSubtitles uploads largely unlicensed. Same copyright lesson as Danish Gigaword's OpenSubtitles (paper 2508.02271). Not openly licensed.
europeana_eu_pl_corpus_raw_speakleash	Aggregated items with mixed per-record rights (PD / CC-BY-NC / rights-reserved). The raw SpeakLeash redistribution is excluded; only a direct rebuild preserving per-record rights metadata may be included.
project_gutenberg_pl_corpus	Only 31 PL books (4.3MB) — PG is ~99% English; Polish PD literature already covered by wolne_lektury (so near-redundant after dedup). Dropped to avoid the PD-in-EU per-work check (PG claims PD-in-US only) for negligible token gain.

Personal & sensitive data

The corpus is built from text that its upstream sources already published under open licences or as official public record. It names public figures acting in a public capacity, for example:

• parliamentary speakers, from PPC (the Polish Parliamentary Corpus) and the Sejm API;
• authorities named in legal acts (EUR-Lex);
• people described in encyclopedic articles (Wikipedia).

The corpus can also hold personal data about private persons, which our filters do not fully remove.

v0.2.6 ships before the NERGAL pass. Only each source's own filters apply (see its datasheet). The NERGAL personal-data pass runs on every source before v0.2.7.

Your rights. If the corpus holds data about you, you can ask us to:

• tell you what it holds (GDPR Art. 15);
• correct it (Art. 16);
• remove it (Art. 17);
• restrict its use (Art. 18).

You can also object to its processing (Art. 21).

Write to takedown@fabryka.ai. Name the record by its source and record id, by a link, or by a phrase from its text.

• We answer within one month. For complex or numerous requests we may extend this by two months, and we tell you why within the first month.
• We ask for proof of identity only when we cannot otherwise tell that the request is yours.
• You can also complain to the President of the Personal Data Protection Office (UODO).

Removal follows the notice-and-takedown procedure below.

Disclaimer & legal

• Provenance in good faith. Per-source licences are reproduced as documented by the upstream sources, to the best of our knowledge. We make no independent legal warranty about the copyright status of any individual document.
• Publication basis. Each source is published on its own basis, given in its datasheet: a licence, the public domain, the status of an official document or another basis the datasheet states. We do not rely on the text-and-data-mining exception to publish any source.
• Acquisition dates. Datasheets give the date a source was fetched where we recorded it; it is not recorded for every source.
• No ownership claim. This release is a curated, license-reviewed, documented aggregation. We claim no ownership of the underlying texts; rights remain with the original authors/rightsholders under their respective licenses.
• Provided "as is", without warranty of any kind, express or implied. This is not legal advice.
• Your compliance is yours. Downstream users must satisfy the licence of each document themselves, as its license field and its source's datasheet give it, including attribution. See Training models on this dataset.
• Copied text in Wikipedia. The Wikipedia build drops pages that carry plwiki's copyright-violation notice, but it cannot detect a paste without one. In a hand-checked sample of 80 pages, 2 contained text copied or closely paraphrased from a copyrighted source: about 2.5%, with a 95% upper bound of 12.9% for each licence label. Both pages are dropped; others like them remain and are removed on notice (below). Method and sample: data/wikipedia/wikipedia.md.
• Notice-and-takedown. Anyone can ask us to remove a record: a rightsholder, a source, or a person the text is about. Write to takedown@fabryka.ai. Name the record by its source and record id, by a link, or by a phrase from its text, and give the reason.
  • We acknowledge the request within 5 working days.
  • We disable a substantiated request within 7 days. The record leaves main. Its id and text hash go on a blocklist, so no later build brings it back.
  • We then remove it from every earlier public revision on the Hub, including tagged releases. We aim to finish within 30 days, with Hugging Face's help where history has to be rewritten. Old commits and tags stay, but the removed files can no longer be downloaded from them. We cannot reach copies that were downloaded before.
  • We record each removal in takedown/register.md, without personal data.
  • The procedure is in takedown/PROCEDURE.md.

License & attribution

The collection (selection, arrangement and our metadata) is released under CC BY-SA 4.0, to the extent of our rights. Each document keeps the licence or status in its license field; the collection licence does not change it, and it does not cover third-party works in official documents or third-party statements in parliamentary texts. Attribution is due to each upstream source (see the datasheets). A source withdrawn after an objection is recorded in the CHANGELOG. A single record removed on request is listed in takedown/register.md; see Notice-and-takedown above.

Training models on this dataset

CC licenses apply only to uses that require permission under copyright or related rights; they do not restrict uses permitted by statutory exceptions. The ShareAlike condition applies when adapted material is shared publicly, and including a work unchanged in a collection does not by itself create adapted material. In our understanding, and in the view of several commentators, training a model does not by itself make the model adapted material of the training texts, so ShareAlike does not apply to model weights or adapters merely because they were trained on such texts; this has not been tested in court. The assessment may differ where a model or adapter contains and can reproduce protected portions of training texts, or where an output reproduces protected expression from them, literally or in recognisably adapted or translated form; sharing such a model or output then requires a basis (the license of the reproduced text, with its conditions, or a statutory exception), and whether ShareAlike also applies in that case is unsettled. Creative Commons has published guidance for those who choose to comply with CC licenses in all cases, stating that it does not take a position on whether and when copyright applies; it recommends that models and outputs based on ShareAlike material be shared under the same license, while noting that this will almost certainly lead to overcompliance. The Free Art License 1.3 has its own copyleft conditions, for derived works (§ 2.3) and for incorporation into a larger work (§ 4), and their application to models is likewise unsettled. Datasets derived from individual documents (paraphrased, summarised or machine-translated versions) and retrieval datastores containing the texts are outside this assessment: sharing them may amount to sharing the texts or adaptations of them; embedding indexes and new databases built from this dataset may need a separate assessment.

If you want to minimise licensing risk for a model trained on this dataset, you can exclude documents whose license field indicates CC BY-SA or the Free Art License. If you include them and ShareAlike or the Free Art License's copyleft does apply to what you share, it is not enough to share it on terms that merely do not restrict the rights granted: CC BY-SA requires a Creative Commons license with the same License Elements (this version or later) or a BY-SA Compatible License, and the Free Art License requires the same or a compatible license. Attribution obligations attach to the individual licensed texts (creator, copyright and license notices, indication of changes); crediting the dataset (for example "Training data: Polish DynaWord, ") helps only if the link leads to a resource that actually contains the information required for each text. Check what the dataset provides for each document: the license and source fields may not include every required notice (for example, author names are missing for some older sources). License metadata may contain errors; filtering reduces, but does not eliminate, this risk.

Models may reproduce portions of training texts, literally or in recognisably adapted or translated form. Where an output reproduces protected expression from a text and sharing it requires permission, it is subject to that text's license (including attribution); facts, ideas, commonplace phrases without original expression and uses covered by statutory exceptions (such as quotation) are not; shortness alone does not make a fragment unprotected.

Removing a text from the corpus does not remove it from models already trained.

For general-purpose model providers

What a provider of a general-purpose AI model may need for its copyright policy and its public summary of training content (AI Act, Art. 53(1)(c) and (d)). Those obligations rest with the model provider. This card is not a compliance guarantee.

• Name and version: Polish DynaWord v0.2.6, https://huggingface.co/datasets/SlayerLab/polish-dynaword, loaded with revision="v0.2.6".
• Size and acquisition: documents and tokens per source in Sources. Each datasheet under data/<source>/ gives the date the source was fetched where we recorded it; it is not recorded for every source.
• Content: human-written Polish text under open licences (Creative Commons, the Free Art License), public-domain works and official documents (legislation, parliamentary records, public-procurement notices). It contains personal data from public sources (Personal & sensitive data).
• Licence per document: the machine-readable license field. Each datasheet gives the source's basis for publication.
• Text-and-data-mining reservations: we did not check machine-readable reservations (for example in robots.txt) systematically, and the corpus does not record them per document.
• Filters and personal data: the gates in Method and each source's own filters, described in its datasheet. v0.2.6 has not had the NERGAL personal-data pass; it runs on every source before v0.2.7.
• Withdrawn sources and change history: Withdrawn since v0.2.5 and CHANGELOG.md.
• Removal requests: the notice-and-takedown procedure under Disclaimer & legal, with the public register in takedown/register.md.

Reproduce

python3 src/build_dynaword.py --all --speakleash-dir <speakleash_zst_dir> --out .
python3 src/make_docs.py

Results

Corpus phrase frequency (normalized by tokens)

Raw counts and token-normalized shares are regenerated from the current Parquet files with src/pattern_frequency_report.py.

• Total token count (tiktoken proxy): 7,036,938,603

Pattern	Count	Share of all tokens
w roku	457,031	0.0065%
klasyfikacji	169,748	0.0024%
ustawa	664,616	0.0094%
artykuł	3,423,411	0.0486%
parlament	1,284,768	0.0183%
rozporządzenie	1,515,578	0.0215%
w pobliżu	81,258	0.0012%
mieszkańców	337,033	0.0048%
Dz.U.	979,489	0.0139%

Per-source shares

source	pattern	count	share of source tokens
biblioteka_nauki	w roku	24,639	0.00417%
biblioteka_nauki	klasyfikacji	7,192	0.00122%
biblioteka_nauki	ustawa	45,001	0.00762%
biblioteka_nauki	artykuł	84,189	0.01425%
biblioteka_nauki	parlament	29,217	0.00494%
biblioteka_nauki	rozporządzenie	18,331	0.00310%
biblioteka_nauki	w pobliżu	2,783	0.00047%
biblioteka_nauki	mieszkańców	30,702	0.00520%
biblioteka_nauki	Dz.U.	31,913	0.00540%
dziennik_ustaw	w roku	33,096	0.00681%
dziennik_ustaw	klasyfikacji	10,896	0.00224%
dziennik_ustaw	ustawa	77,376	0.01592%
dziennik_ustaw	artykuł	27,773	0.00571%
dziennik_ustaw	parlament	42,338	0.00871%
dziennik_ustaw	rozporządzenie	166,383	0.03423%
dziennik_ustaw	w pobliżu	1,087	0.00022%
dziennik_ustaw	mieszkańców	8,025	0.00165%
dziennik_ustaw	Dz.U.	158	0.00003%
edukacja_medialna_pl	w roku	1	0.00014%
edukacja_medialna_pl	klasyfikacji	1	0.00014%
edukacja_medialna_pl	ustawa	31	0.00435%
edukacja_medialna_pl	artykuł	167	0.02342%
edukacja_medialna_pl	parlament	0	0.00000%
edukacja_medialna_pl	rozporządzenie	0	0.00000%
edukacja_medialna_pl	w pobliżu	0	0.00000%
edukacja_medialna_pl	mieszkańców	10	0.00140%
edukacja_medialna_pl	Dz.U.	0	0.00000%
eltec_pol	w roku	108	0.00050%
eltec_pol	klasyfikacji	1	0.00000%
eltec_pol	ustawa	153	0.00071%
eltec_pol	artykuł	174	0.00081%
eltec_pol	parlament	95	0.00044%
eltec_pol	rozporządzenie	35	0.00016%
eltec_pol	w pobliżu	246	0.00114%
eltec_pol	mieszkańców	214	0.00100%
eltec_pol	Dz.U.	0	0.00000%
eurlex	w roku	39,642	0.00170%
eurlex	klasyfikacji	57,321	0.00246%
eurlex	ustawa	30,329	0.00130%
eurlex	artykuł	1,774,561	0.07606%
eurlex	parlament	779,474	0.03341%
eurlex	rozporządzenie	1,200,341	0.05145%
eurlex	w pobliżu	5,754	0.00025%
eurlex	mieszkańców	9,389	0.00040%
eurlex	Dz.U.	914,962	0.03922%
global_voices	w roku	146	0.00423%
global_voices	klasyfikacji	4	0.00012%
global_voices	ustawa	148	0.00428%
global_voices	artykuł	913	0.02643%
global_voices	parlament	348	0.01008%
global_voices	rozporządzenie	23	0.00067%
global_voices	w pobliżu	90	0.00261%
global_voices	mieszkańców	463	0.01341%
global_voices	Dz.U.	1	0.00003%
govpl	w roku	3,228	0.00473%
govpl	klasyfikacji	514	0.00075%
govpl	ustawa	4,562	0.00668%
govpl	artykuł	1,659	0.00243%
govpl	parlament	3,611	0.00529%
govpl	rozporządzenie	4,024	0.00589%
govpl	w pobliżu	649	0.00095%
govpl	mieszkańców	12,715	0.01862%
govpl	Dz.U.	2,921	0.00428%
nkjp1m	w roku	109	0.00419%
nkjp1m	klasyfikacji	9	0.00035%
nkjp1m	ustawa	79	0.00304%
nkjp1m	artykuł	153	0.00588%
nkjp1m	parlament	216	0.00830%
nkjp1m	rozporządzenie	20	0.00077%
nkjp1m	w pobliżu	77	0.00296%
nkjp1m	mieszkańców	159	0.00611%
nkjp1m	Dz.U.	2	0.00008%
nonsa	w roku	522	0.00338%
nonsa	klasyfikacji	100	0.00065%
nonsa	ustawa	70	0.00045%
nonsa	artykuł	3,009	0.01949%
nonsa	parlament	307	0.00199%
nonsa	rozporządzenie	14	0.00009%
nonsa	w pobliżu	316	0.00205%
nonsa	mieszkańców	1,167	0.00756%
nonsa	Dz.U.	0	0.00000%
open_agh_chemistry_pl	w roku	6	0.00106%
open_agh_chemistry_pl	klasyfikacji	9	0.00160%
open_agh_chemistry_pl	ustawa	0	0.00000%
open_agh_chemistry_pl	artykuł	3	0.00053%
open_agh_chemistry_pl	parlament	0	0.00000%
open_agh_chemistry_pl	rozporządzenie	0	0.00000%
open_agh_chemistry_pl	w pobliżu	17	0.00301%
open_agh_chemistry_pl	mieszkańców	0	0.00000%
open_agh_chemistry_pl	Dz.U.	0	0.00000%
open_icm_pl	w roku	19	0.00430%
open_icm_pl	klasyfikacji	9	0.00204%
open_icm_pl	ustawa	30	0.00680%
open_icm_pl	artykuł	121	0.02741%
open_icm_pl	parlament	34	0.00770%
open_icm_pl	rozporządzenie	4	0.00091%
open_icm_pl	w pobliżu	0	0.00000%
open_icm_pl	mieszkańców	27	0.00612%
open_icm_pl	Dz.U.	14	0.00317%
parliamentary	w roku	174,916	0.01163%
parliamentary	klasyfikacji	10,872	0.00072%
parliamentary	ustawa	444,839	0.02957%
parliamentary	artykuł	176,400	0.01173%
parliamentary	parlament	297,729	0.01979%
parliamentary	rozporządzenie	104,305	0.00693%
parliamentary	w pobliżu	4,584	0.00030%
parliamentary	mieszkańców	73,641	0.00490%
parliamentary	Dz.U.	23,615	0.00157%
plwiki_talk	w roku	2,632	0.00474%
plwiki_talk	klasyfikacji	630	0.00113%
plwiki_talk	ustawa	1,365	0.00246%
plwiki_talk	artykuł	75,408	0.13571%
plwiki_talk	parlament	1,241	0.00223%
plwiki_talk	rozporządzenie	464	0.00084%
plwiki_talk	w pobliżu	421	0.00076%
plwiki_talk	mieszkańców	2,297	0.00413%
plwiki_talk	Dz.U.	289	0.00052%
plwiki_user_talk	w roku	2,439	0.00064%
plwiki_user_talk	klasyfikacji	1,213	0.00032%
plwiki_user_talk	ustawa	1,379	0.00036%
plwiki_user_talk	artykuł	1,211,479	0.31796%
plwiki_user_talk	parlament	2,043	0.00054%
plwiki_user_talk	rozporządzenie	618	0.00016%
plwiki_user_talk	w pobliżu	568	0.00015%
plwiki_user_talk	mieszkańców	2,755	0.00072%
plwiki_user_talk	Dz.U.	234	0.00006%
rock_pollub_pl	w roku	22	0.00187%
rock_pollub_pl	klasyfikacji	150	0.01273%
rock_pollub_pl	ustawa	21	0.00178%
rock_pollub_pl	artykuł	173	0.01469%
rock_pollub_pl	parlament	25	0.00212%
rock_pollub_pl	rozporządzenie	19	0.00161%
rock_pollub_pl	w pobliżu	4	0.00034%
rock_pollub_pl	mieszkańców	3	0.00025%
rock_pollub_pl	Dz.U.	12	0.00102%
sejm_api	w roku	2,002	0.00653%
sejm_api	klasyfikacji	125	0.00041%
sejm_api	ustawa	10,937	0.03570%
sejm_api	artykuł	749	0.00244%
sejm_api	parlament	14,159	0.04621%
sejm_api	rozporządzenie	1,088	0.00355%
sejm_api	w pobliżu	85	0.00028%
sejm_api	mieszkańców	2,689	0.00878%
sejm_api	Dz.U.	4	0.00001%
sejm_api_2011_2022	w roku	2,592	0.00965%
sejm_api_2011_2022	klasyfikacji	50	0.00019%
sejm_api_2011_2022	ustawa	8,934	0.03325%
sejm_api_2011_2022	artykuł	1,079	0.00402%
sejm_api_2011_2022	parlament	11,246	0.04186%
sejm_api_2011_2022	rozporządzenie	689	0.00256%
sejm_api_2011_2022	w pobliżu	71	0.00026%
sejm_api_2011_2022	mieszkańców	2,109	0.00785%
sejm_api_2011_2022	Dz.U.	4	0.00001%
sejm_committee_transcripts	w roku	6,980	0.00623%
sejm_committee_transcripts	klasyfikacji	282	0.00025%
sejm_committee_transcripts	ustawa	19,420	0.01734%
sejm_committee_transcripts	artykuł	13,837	0.01235%
sejm_committee_transcripts	parlament	18,245	0.01629%
sejm_committee_transcripts	rozporządzenie	5,255	0.00469%
sejm_committee_transcripts	w pobliżu	146	0.00013%
sejm_committee_transcripts	mieszkańców	4,129	0.00369%
sejm_committee_transcripts	Dz.U.	73	0.00007%
studia_bas	w roku	87	0.00386%
studia_bas	klasyfikacji	85	0.00377%
studia_bas	ustawa	312	0.01383%
studia_bas	artykuł	272	0.01206%
studia_bas	parlament	154	0.00683%
studia_bas	rozporządzenie	180	0.00798%
studia_bas	w pobliżu	8	0.00035%
studia_bas	mieszkańców	359	0.01591%
studia_bas	Dz.U.	476	0.02110%
ted_pl	w roku	3,586	0.00906%
ted_pl	klasyfikacji	311	0.00079%
ted_pl	ustawa	2,236	0.00565%
ted_pl	artykuł	1,342	0.00339%
ted_pl	parlament	691	0.00175%
ted_pl	rozporządzenie	4,275	0.01080%
ted_pl	w pobliżu	18	0.00005%
ted_pl	mieszkańców	487	0.00123%
ted_pl	Dz.U.	4,202	0.01062%
wiadomosci_statystyczne_pl	w roku	52	0.00345%
wiadomosci_statystyczne_pl	klasyfikacji	122	0.00810%
wiadomosci_statystyczne_pl	ustawa	38	0.00252%
wiadomosci_statystyczne_pl	artykuł	322	0.02138%
wiadomosci_statystyczne_pl	parlament	29	0.00193%
wiadomosci_statystyczne_pl	rozporządzenie	54	0.00359%
wiadomosci_statystyczne_pl	w pobliżu	2	0.00013%
wiadomosci_statystyczne_pl	mieszkańców	227	0.01507%
wiadomosci_statystyczne_pl	Dz.U.	46	0.00305%
wikibooks	w roku	311	0.00204%
wikibooks	klasyfikacji	37	0.00024%
wikibooks	ustawa	164	0.00108%
wikibooks	artykuł	721	0.00474%
wikibooks	parlament	281	0.00185%
wikibooks	rozporządzenie	131	0.00086%
wikibooks	w pobliżu	122	0.00080%
wikibooks	mieszkańców	199	0.00131%
wikibooks	Dz.U.	16	0.00011%
wikinews	w roku	449	0.00409%
wikinews	klasyfikacji	638	0.00582%
wikinews	ustawa	389	0.00355%
wikinews	artykuł	2,462	0.02245%
wikinews	parlament	2,064	0.01882%
wikinews	rozporządzenie	168	0.00153%
wikinews	w pobliżu	455	0.00415%
wikinews	mieszkańców	1,011	0.00922%
wikinews	Dz.U.	27	0.00025%
wikipedia	w roku	157,059	0.01280%
wikipedia	klasyfikacji	79,082	0.00645%
wikipedia	ustawa	15,755	0.00128%
wikipedia	artykuł	43,982	0.00358%
wikipedia	parlament	80,390	0.00655%
wikipedia	rozporządzenie	9,008	0.00073%
wikipedia	w pobliżu	55,953	0.00456%
wikipedia	mieszkańców	179,796	0.01465%
wikipedia	Dz.U.	515	0.00004%
wikivoyage	w roku	460	0.00327%
wikivoyage	klasyfikacji	15	0.00011%
wikivoyage	ustawa	29	0.00021%
wikivoyage	artykuł	616	0.00437%
wikivoyage	parlament	154	0.00109%
wikivoyage	rozporządzenie	31	0.00022%
wikivoyage	w pobliżu	5,865	0.04164%
wikivoyage	mieszkańców	2,883	0.02047%
wikivoyage	Dz.U.	0	0.00000%
wiktionary_examples	w roku	7	0.00338%
wiktionary_examples	klasyfikacji	1	0.00048%
wiktionary_examples	ustawa	3	0.00145%
wiktionary_examples	artykuł	12	0.00579%
wiktionary_examples	parlament	1	0.00048%
wiktionary_examples	rozporządzenie	1	0.00048%
wiktionary_examples	w pobliżu	6	0.00290%
wiktionary_examples	mieszkańców	4	0.00193%
wiktionary_examples	Dz.U.	0	0.00000%
wolne_lektury	w roku	1,921	0.00209%
wolne_lektury	klasyfikacji	79	0.00009%
wolne_lektury	ustawa	1,016	0.00111%
wolne_lektury	artykuł	1,835	0.00200%
wolne_lektury	parlament	676	0.00074%
wolne_lektury	rozporządzenie	117	0.00013%
wolne_lektury	w pobliżu	1,931	0.00211%
wolne_lektury	mieszkańców	1,573	0.00172%
wolne_lektury	Dz.U.	5	0.00001%