Datasets:
feat(data): add biblioteka_nauki source (42,071 docs, per-document license + attribution)
Adds biblioteka_nauki as a prebuilt source: 42,071 documents / 1,748,333,935 chars / ~673.5M tokens (tiktoken cl100k proxy).
Why prebuilt
The dataset card notes the live bibliotekanauki.pl probe currently returns HTTP 500 with zero records. This PR removes the dependency on the live upstream: the parquet ships ready-built, and the recovery scripts (recover_licenses.py, recover_attribution.py) are vendored alongside so the build is reproducible when upstream is back.
License gate (per-document, traceable)
The SpeakLeash blob has no per-document license and a large share of upstream is all-rights-reserved, so a per-item gate was mandatory. Licenses were recovered per document from the bibliotekanauki.pl API (/api/{type}/{id} → license.iconId). Only openly-licensed documents are kept:
| license | documents |
|---|---|
| CC BY 4.0 | 28,335 |
| CC BY-SA 4.0 | 13,733 |
| Public Domain | 3 |
All-rights-reserved, NC/ND and unknown are gated out — the parquet license column provably contains only the three values above.
Attribution
Per-document attribution (author · year · title · publisher) re-fetched 1:1 from upstream (43,576/43,576 ok), carried in the parquet attribution column — satisfies the CC BY / CC BY-SA attribution requirement.
Filters (per Dynaword guidelines, minimal)
read 43,576 → kept 42,071 (non-Polish 925, exact dedup 180, OCR alpha-ratio 400, short 0). Details in biblioteka_nauki.md.
Schema note
license + attribution ride as parquet columns. If the repo prefers the strict 6-column schema, happy to move them to a sidecar attribution.jsonl — say the word and I'll update the PR.
Will update README.md