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
102
polish-dynaword / data /nkjp1m /nkjp1m.md
1am's picture
1am
Update dates
b985dee verified
preview code
|
Raw History Blame
2.35 kB

nkjp1m

The manually annotated 1-million word subcorpus of the National Corpus of Polish

Dataset description

  • Source (upstream): https://clip.ipipan.waw.pl/NationalCorpusOfPolish
  • Domain: mixed (balanced reference corpus)
  • Language: Polish (pl)
  • License: CC-BY
  • Created (range): 1957-2008 (recorded for 6,581 of 18,381 documents; per-document in the created column)
  • Added: 2026-06-15

Licensing — traceable basis

The NKJP download page (clip.ipipan.waw.pl) states the manually annotated 1-million-word subcorpus (Podkorpus Milionowy NKJP 1.2) is available under a CC-BY licence; attribution to the NKJP Consortium (IPI PAN) preserved. The page states 'CC-BY' with no version, normalized to CC-BY-4.0 for SPDX. The CC-BY licence covers the corpus text; NKJP's separate tools are GPL and are not included here.

Provenance

Downloaded the IPI PAN tarball (NKJP-PodkorpusMilionowy-1.2.tar.gz) directly by src/fetch_njkp.py, one row per TEI

passage (its paragraphs joined by newlines, ellipsis markers stripped). The same minimal gates as build_dynaword.py are applied per passage (drop < 200 chars, drop non-Polish by diacritic ratio, exact sha1 dedup; the OCR gate is inapplicable). Per-document 'created' is the sample's TEI publication date () where the header records one. Normalized into the DynaWord schema; cl100k token counts via tiktoken (encode_ordinary), and the stats are recomputed from the written parquet.

Statistics

documents characters tokens (tiktoken proxy)
18,381 6,824,905 2,601,201

Per-document license metadata

license documents
CC-BY 18,381

Author metadata present for 18,381 documents. Empty values mean the upstream record did not expose a machine-readable author field.

Statistics were recomputed directly from the released parquet file.

Filters applied (build_dynaword.py)

Minimal, per Dynaword guidelines (heavy filtering left to downstream use):

  • drop documents < 200 chars: 92
  • drop non-Polish (diacritic ratio): 2
  • exact cross-source dedup (sha1): 9
  • OCR alpha-ratio < 0.70 (OCR sources only): 0
  • read 18,484 → kept 18,381

Token counts are a fast tiktoken (cl100k) proxy.