Datasets:

SlayerLab
/

polish-dynaword

Wiktionary contributors license their text under CC BY-SA 4.0: the Wikimedia Terms of Use (read 2026-10-05) and the pl.wiktionary footer, which links creativecommons.org/licenses/by-sa/4.0/deed.pl. Text contributed before June 2023 was under CC BY-SA 3.0, which allows an adaptation under a later version (3.0 §4(b)). v0.2.5 labelled the source CC-BY-SA-3.0; this release labels it CC-BY-SA-4.0.

Not every example sentence is contributor text. Many are quotations from books, press, the NKJP corpus, the web or Wikipedia, used on Wiktionary under the quotation right (Art. 29 of the Polish copyright act). That right does not pass to a corpus, and the CC BY-SA grant does not cover them. The fetcher drops them:

line	lines	released
carries `{{źródło}}` or a `<ref>`	32,664	no
names its source by hand, e.g. "(B. Prus: Lalka)", "(z Internetu)", "(Wikipedia)", "(Mt 4:19)", or is an excerpt marked "(…)" / "[…]"	319	no
shorter than 10 characters	50	no
contributor's own example	45,212	yes, if its entry reaches 200 characters

The hand-attribution patterns come from a survey of the dump and also drop a few lines that end in a Latin gloss in parentheses. Quotations pasted with no source marker cannot be detected and may remain.

To measure them, 100 random released lines (of 7,749; random.seed(20261005)) were searched on the web as exact phrases on 2026-10-05:

result	lines
found only on Wiktionary or sites that republish it	44
not found	41
too short or common for a match to mean copying	13
found elsewhere, copied from Wiktionary (2017 forum post; on Wiktionary since 2007)	1
paraphrase of a 2014 news quote, not verbatim	1
copied from another source	0

0 of 100 puts the share of unmarked quotations below 3.7% (95%, Wilson). Printed or paywalled sources that search engines do not index cannot match. Per line: wiktionary_examples.quotation-sample.jsonl. The source is kept (decided 2026-10-05).

Evidence: wiktionary_examples.license-evidence.json.

Attribution

Each row's author is Wikisłownik contributors, <page URL>, e.g. https://pl.wiktionary.org/wiki/czyta%C4%87. The page history names the contributors; CC BY-SA 4.0 §3(a)(2) allows attribution by a URI. The row id is wiktionary_examples_<page id>.

Provenance

Parsed from the official Wikimedia dump plwiktionary-20261001-pages-articles.xml.bz2 (sha1 a097f6e2b005d1235936a19ed8e800e29a6f2fad) by src/fetch_wiktionary_examples.py; not routed through SpeakLeash. One document per Polish-language headword: the lines of its {{przykłady}} section, in source order. The fetcher also writes wiktionary_examples.citations.jsonl (not part of the corpus) listing, per entry, the sources of the dropped lines.

869,406 pages scanned → 116,157 Polish entries → 31,094 with at least one kept example → 1,829 at 200 characters or more → 1,818 after deduplication.

Statistics

documents	characters	tokens (tiktoken proxy)
1,818	534,521	207,181

v0.2.5 had 8,639 documents (2,889,761 characters). Most of the drop comes from the 200-character document gate: removing the quotations leaves most entries with one or two short examples. The gate is kept (decided 2026-10-05).

Per-document license metadata

license	documents
`CC-BY-SA-4.0`	1,818

Author metadata present for 1,818 documents.

Filters applied (build_dynaword.py)

drop documents < 200 chars: 0 (the fetcher applies the same gate: 29,265 entries)
drop non-Polish (diacritic ratio): 0
exact cross-source dedup (sha1): 11
OCR alpha-ratio < 0.70 (OCR sources only): 0
read 1,829 → kept 1,818

Token counts are a fast tiktoken (cl100k) proxy (~1% off Llama-3); the canonical Llama-3 count is computed at release.

Not checked

quotations pasted without any source marker beyond the 100-line web sample above
whether an example paraphrases a protected text
the 3.0 → 4.0 relabel of pre-2023 text (reading of 3.0 §4(b), not a lawyer's opinion)
personal data in example sentences (the NERGAL scrub runs before release)