SlayerLab/polish-dynaword

Polish DynaWord

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

v0.2.3 stable · 2,710,974 documents · 6.88B tokens (6,881,277,362 tiktoken cl100k-proxy tokens) · 16 sources Updated: 2026-07-26

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 v0.2.3 stable parquets.

Europeana validation artifact (2026-07-10) · first reproducible direct-ingestion shard: 1,000 API records → 810 quality-gated parquet documents, each preserving per-record rights and creator metadata. Stored separately under the legacy path previews/v0.3.1/europeana.parquet; it is not a release of the full dataset and does not alter stable totals.

Releases

Stable releases

version	status	documents	tokens	notes
v0.2.3	stable release	2,710,974	6.88B	Adds community-contributed HPLT v3 Polish WDS 10+9, Global Voices Polish and the NKJP manually annotated 1M-word subcorpus.
v0.2.2	previous stable	2,579,963	6.36B	Added community-contributed govpl: 88,190 docs / 80.7M tokens fetched directly across 133 ministry/agency subsites.
v0.2.1	previous stable	2,491,773	6.28B	12-source stable corpus with license and author metadata columns; adds 1000_novels.
v0.2.0	previous stable	2,490,773	6.22B	Provenance-first corpus from 11 open/official sources.

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

v0.2.3 — current 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. Result: 13 sources, 2,579,963 documents and 6.36B proxy tokens.

v0.2.1 — previous stable

Introduced the canonical eight-column release schema: id, text, source, added, created, token_count, license, author. Added the CC-BY-4.0 1000_novels source 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. This version established the per-source datasheet, traceable-license and exclusion-policy workflow.

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 a workflow and validation track, 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 validates the direct-ingestion contract without changing stable-release totals. v0.3.1-preview remains only as the legacy storage-path label.

What this dataset contributes

The raw texts come from existing open corpora (redistributed via SpeakLeash and, where applicable, fetched from upstream). The value added here is the curation, not the bytes, following the Dynaword methodology:

1. License review per source — each source vetted for an openly-licensed, traceable legal basis (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 the upstream sources and to SpeakLeash as the redistributing aggregator; 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.	v0.2.2, PR #9
Arkadiusz Słota	Built european_hplt_v3_pl: HPLT v3 WDS 10+9 cleaning pipeline, 110,590-document artifact, validation and documentation.	v0.2.3, PR #5
Dawid Majewski	Added the reviewed Global Voices Polish corpus with per-document author attribution and documentation.	v0.2.3, PR #7
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

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 with a documented legal basis (see each datasheet's "traceable basis"), not a vague "public domain".
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)	CC-BY-4.0	243,060	2,378.1M
parliamentary	Polish Parliamentary Corpus (Sejm/Senat)	public-domain (official documents)	324,622	1,646.8M
wikisource	Polish Wikisource	CC-BY-SA-3.0	632,005	801.9M
wikipedia	Polish Wikipedia	CC-BY-SA-3.0	1,171,897	707.2M
european_hplt_v3_pl	HPLT v3.0 Polish — top WDS quality bins 10+9	CC0-1.0	110,590	512.4M
dziennik_ustaw	Dziennik Ustaw + Monitor Polski (Polish primary legislation)	public-domain (official documents)	35,442	486.1M
wolne_lektury	Wolne Lektury (school readings)	CC-BY-SA-4.0 / Wolna Sztuka 1.3	6,141	103.0M
govpl	gov.pl — Polish government press releases	CC-BY-SA-4.0	88,190	80.7M
1000_novels	1000 Novels Corpus (CLARIN-PL)	CC-BY-4.0	1,000	60.5M
wikiquote	Polish Wikiquote (quotations)	CC-BY-SA-3.0	30,363	31.9M
eltec_pol	ELTeC-pol (European Literary Text Collection, Polish)	CC-BY-4.0	100	21.5M
wikivoyage	Polish Wikivoyage (travel guides)	CC-BY-SA-3.0	13,645	17.1M
wikibooks	Polish Wikibooks (open textbooks)	CC-BY-SA-3.0	9,112	15.6M
wikinews	Polish Wikinews	CC-BY-2.5	24,386	12.1M
global_voices	Global Voices Polish	CC-BY-3.0	2,040	3.7M
nkjp1m	NKJP manually annotated 1-million-word subcorpus	CC-BY	18,381	2.6M
total			2,710,974	6,881.3M

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.

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: prepared in the source registry as a direct-upstream rebuild target with per-document license and author metadata. A live probe currently returns HTTP 500 and zero records, so it is not included in v0.2.1 or the v0.3 preview candidate data.
• europeana: prepared in the source registry as a direct-upstream rebuild target with per-record rights statements and creator metadata. A 200-record live direct-rebuild audit found 197 Public Domain Mark records and three CC-BY-SA-4.0 records; all passed the blocking schema, license, duplicate and evaluation-overlap gates. It remains candidate-only until a complete rebuild, date split, and legal review are complete. Raw SpeakLeash Europeana remains excluded.
• Local SpeakLeash shard audit: 12 reviewed source shards were sampled directly from the live .jsonl.zst inputs and passed the blocking gates; four shards are explicitly excluded (open_subtitles_corpus, raw Europeana, 1000_novels_corpus_CLARIN-PL, project_gutenberg_pl_corpus) and two talk shards remain on hold pending a separate source policy. The audit mechanism is aligned with slayer-dsbench.
• Europeana release policy: split conservatively at pre-1929 records for US-sensitive downstream reuse, and keep later/unknown records separately labeled or held until legal review.
• european_hplt_v3_pl: WDS bins 10+9 are included in v0.2.3 after contract validation. HPLT packaging is CC0, while underlying crawled web documents can carry independent rights; downstream near-dedup and web-content review remain recommended before training.

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 + wikisource (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

This corpus contains only text that its upstream sources already published under open licenses or as official public-domain record. It therefore includes names and statements of public figures acting in a public capacity — e.g. parliamentary speakers (PPC), authorities named in legal acts (EUR-Lex), and people described in encyclopedic articles (Wikipedia/Wikisource). No private, non-public personal data was collected or added. If you are a data subject and want content concerning you removed, contact k.wikiel@gmail.com — it will be dropped from the next version (see retroactive-removal policy below).

Disclaimer & legal

• Provenance in good faith. Per-source licenses are reproduced as documented by the upstream sources and by SpeakLeash (the intermediate aggregator), to the best of our knowledge. We make no independent legal warranty about the copyright status of any individual document.
• 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 each upstream license themselves — in particular CC-BY-SA-4.0 attribution and share-alike for derivatives of this dataset, and attribution to the upstream sources and to SpeakLeash.
• Notice-and-takedown. Any source or rightsholder raising a substantiated objection can have material removed: contact k.wikiel@gmail.com; it is dropped from the next version and recorded in the CHANGELOG. Removal is retroactive going-forward (prior immutable snapshots/commits may persist).

License & attribution

Released under CC-BY-SA-4.0 (copyleft inherited from CC-BY-SA sources such as Wikipedia/Wikisource/Wolne Lektury). Attribution due to each upstream (see datasheets) and to SpeakLeash as the intermediate aggregator. Retroactive-removal policy: a source that raises an objection is dropped from subsequent versions, recorded in the CHANGELOG.

Reproduce

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

Results

The corpus-level phrase-frequency recount for the 16-source v0.2.3 release is pending. The previous 15-source tables were removed to avoid presenting stale analysis as current. Release document/token totals above are summed from the validated per-source parquet statistics.