Datasets:

---

SlayerLab

/

polish-dynaword

Like 26

Follow

Fabryka AI 97

eurlex: per-record EUR-Lex links and attribution sidecar (v2#108)

Adds data/eurlex/eurlex.attribution.jsonl: one JSON line per row of eurlex.parquet (242,488), in row order, joined by id. The Parquet and its text are unchanged (sha256 bc7d865e...eb87d).

Each line has a per-record EUR-Lex link built from the identifier at the start of the text, the issuing institution where the text names it, the notes the EUR-Lex legal notice asks for (not the authentic version, converted to plain text), and a ready-made attribution string. Fields: id, text_sha1, source_url, url_from, celex, oj, institution, institution_from, license_url, notes, attribution. They are documented in the card.

• Links: 237,588 rows (213,579 exact, 24,009 to the document or ruling the text belongs to). No link for 4,900 rows (none_<reason> in url_from): 3,419 whose text gives no usable identifier, and 1,481 whose built CELEX number the Publications Office database does not know (none_celex_unknown; source_url and celex are null and the note no_document_link is set; the institution is kept). Those 1,252 numbers are listed in data/eurlex/eurlex.celex-unknown.txt, pinned by SHA-256 in the builder.
• Consolidated texts (17,173 rows starting TEKST skonsolidowany) keep the label LicenseRef-EU-Reuse-Policy; their lines carry the CC BY 4.0 link and the modification note.
• 132 European Parliament minutes that point to roll-call results available on the Parliament website get their EUR-Lex link and a note that those results are not included.
• Institution: 220,297 rows; 22,191 are null by design (corrigenda, budget statements, treaties and agreements, calls and selection notices, texts with no single author). It is a heuristic over the text; a 561-row sample compared with the authors the Publications Office records agrees on 548.

Checks

Manual check: on 2026-10-20 the maintainer opened 5 addresses from the sidecar in a browser (no automated request) and each showed the intended document: an Official Journal L page (Commission decision 2004/898/WE), an Official Journal CE page (European Parliament resolution of 24.3.2009), a CELEX: address for an Advocate General opinion (C-572/10), one for a Commission proposal (COM(2018) 165) and one for a consolidated regulation (1749/96).

Automated check: the addresses were not fetched by script. eur-lex.europa.eu answers automated requests with a bot check (HTTP 202, x-amzn-waf-action: challenge), so work stopped there and nothing was bypassed. Instead every CELEX identifier was checked against the Publications Office Cellar SPARQL endpoint (a different host; its robots.txt allows /webapi/), 517 sequential requests at least 3 s apart, 2026-10-19/20. Identifiers the database does not know are left without a link (see above).

"Found" means a work with that CELEX exists in the Publications Office database, not that EUR-Lex shows this exact text. Not established: beyond the 5 addresses above no page was opened; the Official Journal locator form beyond the two hand-opened addresses (the database confirms only some first pages, none before 2010, none in series CE); that court notices have a CELEX of their own (773 case-law identifiers are not in the database, 769 of them exist there under another descriptor, i.e. as the ruling). Details are in eurlex.license-evidence.json (attribution_sidecar).

Files

• data/eurlex/eurlex.attribution.jsonl (about 187 MB, 178 MiB, LFS)
• data/eurlex/eurlex.celex-unknown.txt (1,252 CELEX numbers the database does not know)
• data/eurlex/eurlex.license-evidence.json, data/eurlex/eurlex.md, src/sources.py
• src/build_eurlex_attribution.py, src/test_eurlex_attribution.py

python3 src/build_eurlex_attribution.py --check rebuilds the sidecar from the pinned Parquet and compares it.

🤖 Generated with Claude Code