Tracking: adding the remaining German Commons sources

#2
by KennethEnevoldsen - opened
Danish Foundation Models org

German Commons (coral-nlp/german-commons) has 41 sources, 36.77M documents and 154.56B GPT-2 tokens. It also comes with document-level propella annotations in openeurollm/propella-annotations, so a subset needs no annotation compute β€” only the join by document id, which is what create.py does.

Sorted by licence, with the size German Commons reports. The license column of each document is the authority, not the table below.

CC0 β€” no conditions (β‰ˆ110B tokens)

Source Split Docs Tokens
Deutsches Zeitungsportal news 8.1M 43.9B
Europeana Newspapers news 3.3M 20.7B
ANNO news 1.9M 8.1B
German-PD cultural 124K 49.3B
BLBooks cultural 3.7K 1.0B
Open Legal Data legal 250K 1.9B
BGH legal 77K 293M
BPatG, BVerwG, BGH (20th c.), BFH, BVerfG, BAG, BVerfGE, Bundesrecht legal β€” ~1.0B
Plenarprotokolle / Drucksachen des Bundestages political 4.9K 845M
TEDEUTenders economic 57K 111M

CC-BY 4.0 (attribution, β‰ˆ17B tokens)

Source Split Docs Tokens
MOSEL cultural 3.1M 3.2B
SBB Fulltexts cultural 2.6M 359M
EurLEX legal 65K 201M
One Million Posts web 946K 95M
German Political Speeches political 6.7K 29M
WikiNews news 23K 14M
Wikiquote cultural 8.6K 6.7M

CC-BY-SA 4.0 (share-alike, β‰ˆ6.5B tokens) β€” Wikipedia 2.9B, Wikisource 348M, DiBiLit 216M, Wikibooks 180M, Polytechnisches Journal 50M, Wikivoyage 42M, DiBiPhil 32M, Wikiversity 28M, Wikipedia Discussions 1.2B

Needs a licence ruling before use β€” EuroVoc (EUPL, 2.0B), YouTube Commons (14.5B), The Stack (1.1B), DOAB (167M), OpenAlex (414M), arXiv (103K): per-document Various, would need to be checked document by document.

Suggested first wave, in order of value per unit of work: Zeitungsportal/ANNO (news), German-PD (books), Open Legal Data (legal), Plenarprotokoll + Drucksachen (political), Wikipedia (encyclopedic). Note that the news sources are OCR'd historical newspapers, so quality filtering matters more there than elsewhere.

Danish Foundation Models org

Status update, since the list above is now mostly being executed:

Added (21 sources, all CC0 / CC-BY / CC-BY-SA at document level): wikipedia, wikidiscussions, wikisource, wikivoyage, wikiquote, wikibooks, wikiversity, wikinews, openlegaldata, plenarprotokolle, drucksachen, germanpoliticalspeeches, mosel, sbbfulltexts, eurlex, onemillionposts, tedeutenders, blbooks, dibilit, dibiphil, polyjournal β€” plus the existing bverfgaes and bundesrecht.

Not added, with reasons:

  • GermanPD β€” the documents are whole books of ~1.2M tokens each; they need chunking before they are usable, and the OCR is visibly poor.
  • Deutsches Zeitungsportal (43.9B), Europeana Newspapers (20.7B), ANNO (8.1B) β€” the largest remaining block and the only route to a news domain, but all three are historical newspaper OCR. ANNO in particular renders Fraktur as garbage (Mr beeKren uns, kiemit an^ΓΌzieiAkn…). If we take them, we should filter on the content_quality / content_integrity annotations rather than ingest 70B tokens as is.
  • YouTube Commons, The Stack, DOAB, OpenAlex, arXiv, EuroVoc β€” per-document Various or EUPL; see the licence discussion.
  • FineWeb-2 and HPLT β€” see the closed discussion: only the collections are openly licensed, not the documents.
Danish Foundation Models org

Backlog state as of 31 October. Every number here is computed from the files in this repository or from a fresh probe of German Commons, not from the collection card.

In the corpus: 32 sources, 20.80M documents, 17.80B tokens β€” recomputed from the per-source descriptive_stats.json. bgh (77,258 decisions of the Federal Court of Justice, 2000-2024, 236.2M tokens) is the newest addition and lands with v0.0.5. By domain, in the same computation: news 6.74B, books 4.11B, encyclopedic 2.38B, legal 2.36B, social media 1.03B, speeches 648.5M, spoken 250.6M, other 146.4M, reports 139.7M.

Building now: cellar (EU law and publications, 1,876,700 rows; 57.7% of the extraction done at 12:20 UTC).

Queued next, nothing needed from you:

  • germanpd β€” 123,592 scanned public-domain books, CC0 on every probed row. Whole books.
  • zeitungsportal β€” 8.1M OCR-ed newspaper pages. The probe read 35,376 documents of the first part, all CC0-1.0; no id carries a date, so created is the documented publication span of the archive.
  • europeana β€” 3.3M OCR-ed newspaper pages. Same shape: 109,212 documents of the first part, all CC0-1.0.
  • reichtagsprotokolle β€” 213 volumes of the Reichstag protocols, CC-BY-SA-4.0 on all 213 rows. Whole volumes ship whole; a mean of 497,582 words per volume is the format, not a defect.
  • wiki-comments, europarl-de β€” ports from the Danish and Dutch Dynawords, same upstream family.
  • ris-at, fedlex-ch β€” Austrian and Swiss law. Official works are outside copyright, but each needs its own bulk download and parser.

Waiting on one code change, not on you: youtube-commons, doab, the-stack. The builder asserts a single licence tuple for a whole source, so a mixed-licence source stops the build. A per-document allowlist turns them into ordinary builds and is being written.

Waiting on a decision from you:

  • eurovoc β€” EUPL-1.2 on every one of the 76,588 rows probed. 246K documents. Recommendation: skip. It is a multilingual thesaurus, so it contributes vocabulary lists rather than running text.
  • openalex β€” mixed CC0 / CC-BY / CC-BY-SA within one source. With the allowlist in place the permissive rows could be kept and the rest dropped and counted; whether that is acceptable for a scholarly-abstract source is your call rather than mine.

A correction to the status comment above. It says the German-PD books "need chunking before they are usable". That is not how this corpus works: the unit is the upstream document, so long books ship whole β€” the corpus already carries documents up to 2.9M tokens, and the mean document length is 856 tokens because most sources are short documents. Poor OCR belongs in the subset card under Limitations, not in a filter, so both the books and the newspaper OCR go in as they are and whoever trains decides how to window and filter them.

The sources outside German Commons (Common Crawl for German, FragDenStaat, municipal council systems, Bundesanzeiger) are scraping projects rather than downloads and stay in #6 until you say otherwise.

Danish Foundation Models org

Coverage of coral-nlp/german-commons as of v0.0.13: 39 of its 41 sources are in the corpus, each as its own subset, 32,638,820 documents and 125.66B tokens over the German Commons material.

The two that are not in need no decision from anyone:

  • subset=Political/source=EuroVoc: dropped on 3 November, every probed document is EUPL-1.2.
  • subset=Scientific/source=arXiv: empty upstream, 0.00 GB in one parquet file with no German documents.

bundesrecht is in the corpus but no longer from the German Commons grain: German Commons keeps only the first norm of each act, so it was rebuilt from the C-DBR release in v0.0.11 (6,132 acts, 51.82M tokens, one document per act).

upstream source subset documents tokens
GermanPD germanpd 123,914 41.82B
Deutsches Zeitungsportal zeitungsportal 8,076,160 36.66B
Europeana Newspapers europeana 359,826 17.42B
Youtube Commons youtubecommons 2,809,714 10.23B
Anno anno 1,910,281 6.73B
SBB Fulltexts sbbfulltexts 3,127,203 2.58B
Wikipedia wikipedia 2,930,224 2.38B
Open Legal Data openlegaldata 249,909 1.51B
Wiki Discussions wikidiscussions 8,349,019 0.96B
The Stack thestack 421,350 0.89B
BLBooks blbooks 3,714 0.81B
Reichtagsprotokolle reichstagsprotokolle 522 0.53B
Drucksachen des Bundestages drucksachen 3,017 0.39B
OpenAlex openalex 47,733 0.33B
Wikisource wikisource 240,689 0.28B
MOSEL mosel 2,605,569 0.25B
Plenarprotokolle des Bundestages plenarprotokolle 1,833 0.24B
Entscheidungen des Bundesgerichtshofs bgh 77,258 0.24B
DiBiLit dibilit 2,062 0.16B
EurLEX eurlex 64,934 0.16B
Entscheidungen des Bundespatentgerichts bpatg 30,705 0.15B
Polytechnisches Journal polyjournal 346 0.14B
Directory of Open Access Books doab 1,939 0.12B
Entscheidungen des Bundesverwaltungsgerichts bverwg 27,185 0.10B
TEDEUTenders tedeutenders 57,213 0.09B
One Million Posts onemillionposts 945,997 0.07B
Entscheidungen des BGH in Strafsachen aus dem 20. Jh. bgh_strafsachen 36,062 0.07B
Entscheidungen des Bundesfinanzhofs bfh 10,885 0.06B
Deutsches Bundesrecht bundesrecht 6,132 0.05B
Wikibooks wikibooks 27,292 0.04B
Entscheidungen des Bundesarbeitsgerichts bag 5,624 0.04B
Wikivoyage wikivoyage 20,370 0.03B
Entscheidungen des Bundesverfassungsgerichts bverfg 8,028 0.03B
DiBiPhil dibiphil 270 0.02B
German Political Speeches germanpoliticalspeeches 6,678 0.02B
Wikiversity wikiversity 16,368 0.02B
Amtliche Entscheidungssammlung des BVerfG bverfgaes 919 0.02B
Wikinews wikinews 23,265 0.01B
Wikiquote wikiquote 8,611 0.01B

Sizes are the built subsets, not the collection card's table (which disagrees with its own parquet for several sources).

Danish Foundation Models org

Two corrections to the coverage note above.

subset=Scientific/source=arXiv is not empty: its part file holds 8 documents and 103,478 tokens, every row German text under a real arXiv id, with the licence mixed per row (CC-BY-4.0 x4, CC-BY-SA-4.0 x3, CC0-1.0 x1). The 0.00 GB was 132 KB rounded down, and "no German documents" was simply wrong.

The 93,689 rows the viewer shows for the scientific config are its six splits together - openalex 47,733, wikibooks 27,292, wikiversity 16,371, doab 1,939, polyjournal 346, arxiv 8 - and five of them are already subsets here at those counts.

Kenneth has decided to drop arXiv, so both misses are decisions rather than gaps: 39 of 41 sources, with eurovoc and arxiv out.

One nit on the collection card's own table: wikibooks and polyjournal look swapped (it lists wikibooks as 346 documents and polyjournal as 27,292; the parquet has polyjournal 346 and wikibooks 27,292).

Sign up or log in to comment