Datasets:

---

SlayerLab

/

polish-dynaword

Like 29

Follow

Fabryka AI 101

Holds uzp_orzeczenia_pl off main until v0.2.7 (v2#112). The source was never released.

Why. The G1 pass (Hub PR #193) still leaves part of the sole traders' firm names and addresses unmasked. Round 2 of its rules reached 92.2% on firms and 68.7% on addresses on the held-out sample, against a bar of 90% and 85%.

Removed

• the shard and its sample, attribution, decisions, stats, QA, overlap-audit and checksum files;
• the datasheet uzp_orzeczenia_pl.md and the source's NOTICE.md;
• src/test_uzp_orzeczenia_pl_contract.py;
• the source's row in the README plan table, its line in the CHANGELOG plan, and its section in the root NOTICE.md.

Kept for v0.2.7

• the registry entry (release: None), with its legal note updated;
• src/build_uzp_orzeczenia_pl.py and src/test_build_uzp_orzeczenia_pl.py;
• artifacts/uzp_orzeczenia_pl_ontology_manifest.json;
• the legitimate-interest note, moved to artifacts/uzp_orzeczenia_pl_legitimate_interest.md, since data/ holds only sources with a datasheet.

The README and CHANGELOG plan totals become 14 sources, 1,250,697 documents and 834,746,134 tokens. AGENTS.md and artifacts/source_findings.md record the hold. The dataset config reads data/*/*.parquet, so no config changes.

Not in this PR. Older revisions of main still serve the shard until its LFS object is deleted; that is a separate step.

Checks. python3 -m pytest src/ in an LFS-pointer checkout: 129 failed / 1931 passed / 22 skipped / 108 errors. Clean main gives 130 / 1937 / 23 / 108 in the same checkout. The failure lists differ by one test fewer (test_shard_carries_the_registry_date[uzp_orzeczenia_pl]); the passes and skips that go are the uzp contract test and uzp's parameters in per-source tests. No new failure.

🤖 Generated with Claude Code