Datasets:

---

SlayerLab

/

polish-dynaword

like 13

Follow

Slayer Labs 28

Docs-only. No data shards, no sources.py changes.

artifacts/source_findings.md (new) — append-only log of sources already investigated, version-independent (unlike the *_v0_3.md point-in-time snapshots). Each entry records verdict, enumeration method, license evidence (self-verified vs second-hand), blockers, and the lesson. Current entries: zpe.gov.pl, ParlaMint-PL, Sejm API interpelacje, CLARIN-PL DSpace/PELCRA, NIK, a billion-scale survey, and FineWeb-2/FineWeb2-HQ with a quality_score-to-token-yield table.

The point is that a rejection is as useful as an acceptance — nobody should spend a day re-discovering that a portal renders its license footer in JS.

AGENTS.md — three additions:

• Check the findings log before starting, and append to it either way.
• What the PR must show: a verbatim data sample inline in the PR, and provenance/licensing written into the datasheet — concrete origin, the exact page or statute the license claim rests on, what the texts are, how they were filtered, and the argument for inclusion.
• Normalize the text before you build: build_dynaword.py applies only minimal gates and normalize_schema.py never touches text, so normalization is the fetcher's job — Unicode, whitespace, structural junk, numbering (with the caveat that in dziennik_ustaw/saos article numbering is content), boilerplate by frequency, and PII. Two rules: the normalization must live in the committed script, and it must be reported in the datasheet.