kacperwikiel/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).

v0.2.0 stable · 2,490,773 documents · 6.22B tokens (tiktoken proxy; canonical Llama-3 count at release) · 11 sources

v0.3.0-preview in progress · quality/diversity remix workflow, legal-style downweighting, and filtered contemporary Polish web candidates. Current local HPLT candidate run: 565,827 documents / 496.5M tokens from ashtok897/european-hplt-v1; not yet merged into the stable parquet release.

Versions

version	status	documents	tokens	notes
v0.2.0	stable release	2,490,773	6.22B	Provenance-first corpus from 11 open/official sources.
v0.3.0-preview	workflow + candidate data in progress	+565,827 candidate docs	+496.5M candidate tokens	Filtered European HPLT Polish web candidate, generated locally with chunked/parallel filtering; pending dedup, QA, and final mix weighting.

If the current HPLT candidate is merged as-is, the working corpus would be approximately 3,056,600 documents / 6.72B tokens before dedup and mix reweighting. The final v0.3 release is expected to be a training mix, not a raw append: legal/parliamentary/official-document sources should be capped and sampled rather than allowed to dominate by raw token count.

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.
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).

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
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
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
total			2,490,773	6,221.4M

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.

v0.3 quality roadmap and current status

The v0.2 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, 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;
• treat GPT-2 124M as a cheap dataset probe, not proof of final model quality.

Primary v0.3 web candidate: ashtok897/european-hplt-v1. Its card reports Polish pl coverage of 1,891,358 documents and ~1.98B estimated tokens with HPLT WDS quality scores, language confidence, URL provenance, MIME type, and web-register metadata. The helper src/filter_european_hplt.py streams this dataset and filters for Polish, non-machine-translated, non-boilerplate, domain-filtered natural web text. The dataset card marks it CC0-1.0, inherited from HPLT v3, but because it is web-crawl derived it remains subject to source/provenance review before a final release.

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%.

Current HPLT filtering status (2026-06-16):

• input: ashtok897/european-hplt-v1, actual HF train parquet files split across workers;
• filter output generated locally under data/european_hplt_pl_parallel_500m_v2/;
• kept 565,827 documents and 496,457,232 estimated tokens;
• compressed parquet output size: about 1.0 GB;
• this is a candidate source snapshot, not yet a stable release artifact.

Reproduce the chunked/parallel candidate run:

python3 src/launch_hplt_parallel.py \
  --workers 4 \
  --out-dir data/european_hplt_pl_parallel_500m_v2 \
  --screen-prefix hplt2 \
  --target-total-tokens 500000000

Monitor progress:

screen -ls
tail -f logs/filter_european_hplt_hplt2_w0.log
find data/european_hplt_pl_parallel_500m_v2 -maxdepth 2 -name '*.parquet' -ls

TVP Info-derived news data is currently blocked: the HF upload WiktorS/polish-news has an apache-2.0 tag, but its README says the articles were obtained from tvp.info.pl, and no upstream TVP Info open license has been verified. It should only be included with explicit permission or authoritative upstream open-license evidence.

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	Aggregated items with mixed per-record rights (PD / CC-BY-NC / rights-reserved). Needs per-record edm:rights filter before any inclusion.
1000_novels_corpus_CLARIN-PL	CC-BY-4.0 label, but 'novels' likely include in-copyright contemporary works; verify titles/years on CLARIN handle 11321/312 before inclusion.
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

Corpus phrase frequency (normalized by tokens)

To show how frequent legal and discourse markers are across the corpus, we can report counts normalized by token count per source and globally. Raw counts + percentages are generated from the current parquet data and source token counts:

python3 src/pattern_frequency_report.py --data-root . \
  --out-md pattern_frequency_report.md \
  --out-hf artifacts/pattern_frequency_hf_snippet.md \
  --out-png artifacts/pattern_frequency.png

pattern_frequency_report.md contains full source-by-source breakdown. artifacts/pattern_frequency_hf_snippet.md is the exact block for Hugging Face model card.

pattern	count	share of all corpus tokens
w roku	434,882	0.0070%
klasyfikacji	129,963	0.0021%
ustawa	586,803	0.0094%
artykuł	2,035,630	0.0327%
parlament	1,201,401	0.0193%
rozporządzenie	1,490,399	0.0240%
w pobliżu	77,561	0.0012%
mieszkańców	240,332	0.0039%
Dz.U.	939,966	0.0151%

Per-source normalized shares:

• w roku
• klasyfikacji
• ustawa
• artykuł
• parlament
• rozporządzenie
• w pobliżu
• mieszkańców
• Dz.U.

Hugging Face Model Card block

Wklej dokładnie artifacts/pattern_frequency_hf_snippet.md do sekcji Results w model card (README.md repozytorium na HF).