Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 79

Summary

Adds nkjp1m, the manually annotated 1-million-word subcorpus of the National Corpus of Polish (Podkorpus Milionowy NKJP 1.2, IPI PAN) as a new source. It is a mixed-domain balanced reference corpus of Polish text, released under CC-BY.

Dataset

• Artifact: data/nkjp1m/nkjp1m.parquet
• Documents: 18,381 (one row per TEI <div> passage)
• Characters: 6,824,905
• Tokens (tiktoken cl100k proxy): 2,601,201
• Domain: mixed (balanced reference corpus)
• Language: Polish (pl)
• License: CC-BY (upstream), normalized to CC-BY-4.0 for SPDX
• Schema: DynaWord standard 8 fields (id, text, source, added, created, token_count, license, author)

Provenance and cleaning

Downloaded the IPI PAN tarball (NKJP-PodkorpusMilionowy-1.2.tar.gz) via src/fetch_njkp.py. Emits one row per TEI <div> passage, with its <ab>x paragraphs joined by newlines and ellipsis markers stripped. Each <div> is a contiguous excerpt from one source document and kept as a separate row.

The same minimal gates as build_dynaword.py are applied per passage (heavy filtering is left to downstream use): min-length (< 200 chars), non-Polish diacritic ratio, exact cross-source dedup (sha1). Read 18,484 -> kept 18,381. Token counts use tiktoken cl100k, and stats are recomputed from the written parquet.

Licensing and attribution

• The NKJP download page (clip.ipipan.waw.pl) states the manually annotated subcorpus is available under a CC-BY licence (no version stated, normalized to CC-BY-4.0 for SPDX). NKJP's separate tools are GPL and are not included here.
• Attribution to the NKJP Consortium (IPI PAN) is preserved.
• Upstream: https://clip.ipipan.waw.pl/NationalCorpusOfPolish

Limitations

• No version bump: this PR adds the source and its data/docs but does not bump the dataset version. The README phrase-frequency tables and figures are not refreshed and still exclude nkjp1m; re-run src/pattern_frequency_report.py to refresh them.
• created dates are only available for 6,581 of 18,381 documents (those whose TEI header records a publication date).

Files added

• data/nkjp1m/nkjp1m.parquet - the corpus shard (18,381 rows).
• data/nkjp1m/nkjp1m.stats.json - read/kept/gate counts and totals.
• data/nkjp1m/nkjp1m.md - source documentation card.
• src/fetch_njkp.py - fetch + build script.

Files modified

• src/sources.py - registers the nkjp1m source entry.