artifacts/bartek_source_ingestion_plan_2026-07-02.md

Bartek source ingestion plan

Prepared from source-expansion work by Bart Kobyliński for Biblioteka Nauki and Europeana.

Decision

Accept the two extra columns in the main parquet schema:

id, text, source, added, created, token_count, license, author

Reason: license and author are document-level provenance fields. Keeping them inside the parquet means they survive shuffling, filtering, deduplication, and training-mix extraction. A separate sidecar file is useful for audit exports, but it should not be the only place where rights metadata lives.

Contract for the PR

Bartek should provide:

• normalized parquet/jsonl.zst files for biblioteka_nauki and europeana, or a script that regenerates them from the upstream services;
• per-record license and author fields where upstream metadata exists;
• per-source stats: documents, characters, token count, license distribution, author coverage, dropped-record counts;
• enough source notes to reproduce the exact upstream query/API/interface and the license decision logic.

The project-side scripts now support the expanded schema and will regenerate:

• per-source datasheets in data/<source>/<source>.md;
• the README source table and totals;
• the changelog source list.

Source policy

Biblioteka Nauki

Include as biblioteka_nauki only through the direct upstream rebuild. Do not treat the old SpeakLeash aggregate as sufficient evidence. Each record should carry its upstream license and author/creator metadata when available.

Suggested source description:

Biblioteka Nauki is a Polish academic-text source fetched directly from the upstream site/interface. Records are accepted only when a reusable license is available in the upstream metadata. Per-document license and author metadata are preserved in the corpus.

Europeana

Include as europeana only through the direct upstream rebuild preserving per-record rights statements. The raw SpeakLeash aggregate remains excluded because it is mixed-rights and insufficiently traceable.

Release split:

• created/published < 1929: conservative public release slice for US-sensitive downstream users;
• created/published >= 1929 or unknown date: keep separately labeled, or hold until legal review confirms the release scope.

Dataset-card wording should say that Europeana rights statements are evaluated primarily for Poland/EU and that public-domain status can depend on jurisdiction.

Checklist before merge

• Run the source rebuild or ingest script.
• Validate parquet schema has the eight canonical columns.
• Recompute token counts and stats from the generated files.
• Run python3 src/make_docs.py.
• Inspect the README totals and source datasheets.
• Confirm Europeana is either split at 1929 or explicitly marked as jurisdiction-sensitive.
• Keep raw SpeakLeash Europeana excluded unless per-record rights metadata is preserved.