Datasets:

---

SlayerLab

/

polish-dynaword

Like 24

Follow

Fabryka AI 87

The rock_pollub_pl datasheet, NOTICE and registry entry cited evidence paths that never reached this repository, and the source had no checksums.json. The text was generated for the contributor repo's artifacts/ layout, but the PR flattened only some of those files into data/rock_pollub_pl/.

Commit 1: repair the paths

• rock_pollub_pl.checksums.json (new): SHA-256 of every other file in the directory, in the sibling format. LFS files are listed by LFS oid, and each oid was checked against the real file.
• NOTICE.md: artifacts/attribution.jsonl now points to rock_pollub_pl.attribution.jsonl. All 8 text_sha256 values in it match the parquet.
• rock_pollub_pl.md: the front-matter path now points to the real parquet; it was data/train-00000-of-00001.parquet. A new "Evidence files" section lists what is here and what is not.
• src/test_rock_pollub_contribution.py: two tests are added. One checks that the checksums cover and match every other file. The other checks that every path cited in the datasheet and NOTICE exists; it fails on the old text.

Commit 2: add the contributor inventory; re-verify the PDFs

• rock_pollub_pl.inventory.json and rock_pollub_pl.target_audit.json (new) are byte-identical to artifacts/inventory.json and artifacts/target_audit.json in PiotrSty/rock-pollub-pl-books at v1.0.0 (76f0b7f). Their canonical-JSON SHA-256 equals the ontology manifest's content addresses (33db1ffd…, cddb52e9…), and their raw SHA-256 equals the contributor's own checksums.json.
• The datasheet now gives the 8-of-20 breakdown from the inventory:
  • The collection had 771 archived items. The CC BY-SA 4.0 search returned 743.
  • 136 failed the item-level filter (licence in dc.rights, Polish, Book type, book subtype). The script skips these without a per-item record.
  • The other 607 were inspected: 597 were rejected because the original PDF bitstream does not state CC BY-SA 4.0, 2 because the PDF is over the 50 MiB limit, and 8 were acquired.
• PDF re-verification on 2026-10-07: all 8 PDFs (117,774,897 bytes) were downloaded again from their content_url. Byte count, SHA-256 and the repository's MD5 match for 8 of 8. Each book's imprint page states CC BY-SA 4.0.
• src/sources.py and the registry template in src/fetch_rock_pollub_pl.py now read "URLs, PDF checksums, attribution and per-candidate acquisition rejections are preserved". Commit 1 had dropped "exclusions are preserved" from the registry, and the script still said it. With the inventory present, the narrower wording is true, and both places now say the same.
• Two more tests check that the two files match the manifest pins, and that the inventory backs the datasheet counts.

Note: the inventory includes the full repository metadata of the 8 acquired books, including the authors' public ORCID iDs (11 in total). It contains no email addresses or phone numbers. The rejected entries hold only item id, title and reason.

No data change. README and CHANGELOG are untouched. Not in this repository: the contributor's run.json and checksums.json, the raw PDFs and the extracted texts. The source still has no license-evidence.json.

Tests: full pytest shows no new failures against main 440471b plus commit 1. With the real LFS files, all 6 tests in test_rock_pollub_contribution.py pass. In an LFS-skipped worktree, the parquet- and checksum-dependent tests fail, as their twins in other sources do.

🤖 Generated with Claude Code