Sources outside German Commons: what the other Dynawords have that we don't

#6
by KennethEnevoldsen - opened
Danish Foundation Models org

Mapped against the other Dynawords, this corpus is missing several categories that the Danish, Dutch, Faroese and Swedish ones all cover. None of them come from German Commons, so each needs its own licence check.

Contemporary news. Every other Dynaword has a newspaper or press corpus. German Commons' news domain is historical newspaper OCR (Zeitungsportal 43.9B, Europeana 20.7B, ANNO 8.1B tokens) plus Wikinews. There is no contemporary press, and no openly licensed route to one that I have found — dpa, the regional publishers and the major weeklies are all rights-reserved.

General web. The Dutch corpus has c5_filtered (Common Crawl filtered to Creative Commons pages). German Commons has no crawl at all, and FineWeb-2 / HPLT are not licensed per document. A CC-filtered crawl for German is the single biggest gap.

Government and administrative records. The Dutch corpus has openraadsinformatie (municipal and provincial council documents), officielebekendmakingen and woogle (freedom-of-information disclosures). German equivalents exist but are not aggregated anywhere as an open corpus: FragDenStaat (FOI requests), the municipal Ratsinformationssysteme, the Bundesanzeiger. Each would be a scraping project rather than a download.

European Parliament verbatim reports (German). The Dutch corpus has european_parliament from the EP's plenary minutes under CC-BY. German is a full EP language, so the same source yields a large, clean, permissively licensed corpus of political speech — considerably better than the OCR'd parliamentary material we have.

Public-domain books. The Danish corpus has gutenberg. We now have GermanPD (excluded for now, whole-book documents), BLBooks, SBB Fulltexts, Wikisource, DiBiLit and DiBiPhil, so books are reasonably covered; Project Gutenberg German would add curated, cleanly transcribed literature on top.

Sub-national law. Austrian RIS and Swiss fedlex both publish official works that are not protected by copyright (§ 5 UrhG and its equivalents), with bulk download. Our legal domain is German federal law plus EU law only.

Proposed order, subject to the licence check: EP verbatim reports, then Project Gutenberg German, then RIS/fedlex. I have not added any of them — this is a list to react to, not a plan I have started.

Danish Foundation Models org

Closed by c5_filtered, which went out in v0.0.12 just now: the German part of C5 (the Common Crawl pages that state a Creative Commons licence), seven snapshots from 2019-30 to 2025-05, 3,982,275 documents / 5.93B tokens, one document per page, whole, with the licence on the row - 3.02M CC-BY-SA, 0.82M CC-BY, 0.14M CC0.

Pages that name a non-commercial or no-derivatives licence, cc-unknown, the Public Domain Mark or two different licences at once are dropped; exact-text duplication across the snapshots removed another 1,777,450. Card: https://huggingface.co/datasets/danish-foundation-models/german-dynaword/blob/main/data/c5_filtered/c5_filtered.md

The corpus is now 43 subsets and 138.46B tokens. Contemporary news is still open - I have not found an openly licensed route to one.

Remaining from your list: the EP verbatim reports (discussion #17, behind a bot wall on every document host) and the five sources Vladimir proposed (#20-#24), of which ParlaMint-AT is building now.

Danish Foundation Models org

A source outside German Commons went in with v0.0.21: landesrecht_sachsen, the consolidated law of the Free State of Saxony from REVOSax (https://www.revosax.sachsen.de) - 4,134 Vorschriften, 21.23M tokens, one document per regulation whole, public-domain per row (§ 5 UrhG), id from the portal, null metadata.

That closes state law for one Land. The other fifteen publish on their own portals (mostly juris-hosted: landesrecht-bw.de, gesetze-bayern.de, landesrecht-hamburg.de, ...; NRW on recht.nrw.de, Niedersachsen on voris.wolterskluwer-online.de), each a separate fetcher, so they follow one state at a time rather than as one dump.

Still open in this thread's scope: German EP verbatim (#17, behind the EP bot wall), and the domains the corpus has no source for at all - Subtitles, Readaloud, Conversation - where no openly licensed German bulk source has turned up yet.

Sign up or log in to comment