gmrrh/de-wiki

German Wikipedia (Cleaned)

A cleaned, plain-text version of the German Wikipedia for language model pre-training. Each line is a single article; wiki markup, references, and boilerplate navigation sections have been removed.

Articles	~2,000,000
Size	~10 GB (UTF-8 plain text)
Format	JSONL, one {"text": "..."} object per line
Source dump	dewiki-20260601-pages-articles-multistream.xml.bz2
Source URL	https://dumps.wikimedia.org/dewiki/20260601/
License	GFDL + CC-BY-SA 3.0 (inherited)
Language	German (de)

What's in this dataset

Plain-text article bodies. Section structure (headers, paragraphs) is preserved as line breaks; wiki markup, templates, infoboxes, tables, image captions, magic words, references, and the standard end-of-article navigation sections are stripped.

Every record is a single article. The corpus is suitable for:

• Continued pre-training of base language models on German
• Causal language modeling from scratch
• Masked language modeling (BERT-style)
• Domain-adaptive pre-training for German-heavy applications

What was filtered out

The following pages from the raw MediaWiki dump are excluded:

• Non-article namespaces — Wikipedia:, Portal:, Hilfe:, Kategorie:, Vorlage:, Modul:, MediaWiki: (any namespace != 0)
• Redirects
• Stubs — articles with fewer than 1500 characters after markup stripping
• Oversized articles — articles longer than 80,000 characters after cleanup (typically list-articles and bot-generated entries)
• Disambiguation pages — titles containing "(Begriffsklärung)" or "(disambiguation)"
• List pages — titles starting with "Liste " or "List "

Within retained articles, the following residual cleanup is applied:

• Lines beginning with mini| are removed (leftover cells from stripped Infobox templates)
• Runs of three or more consecutive newlines are collapsed to two

How to use

Load with the datasets library:

from datasets import load_dataset

ds = load_dataset("gmrrh/de-wiki", split="train")

# ds[i]["text"] is the full article body as a single string

# Iterate for training
for ex in ds:
    text = ex["text"]
    # tokenize, pack into chunks, train

Streaming for environments where the ~10 GB download is inconvenient:

from datasets import load_dataset

ds = load_dataset(
    "gmrrh/de-wiki",
    split="train",
    streaming=True,
)
for ex in ds:
    text = ex["text"]
    # ...

Data fields

Field	Type	Description
text	string	Article body, plain text, UTF-8, line-broken by section

There is a single train split. No held-out splits are provided; this is a pre-training corpus, not an evaluation benchmark.

How this dataset was built

The pipeline is reproducible from public sources. Source code for the cleaning scripts is published alongside this dataset.

1. Download the raw dump from https://dumps.wikimedia.org/dewiki/<YYYYMMDD>/dewiki-<YYYYMMDD>-pages-articles-multistream.xml.bz2. The dump used for the current version is 20260601.
2. Stream-parse the MediaWiki XML with mwxml (forward-only iterparse, no full XML tree in memory).
3. Drop pages by namespace, redirect status, and title patterns before any expensive text processing.
4. Strip wiki markup with mwparserfromhell.parse(text).strip_code(normalize=True, collapse=True), which removes templates, tables, wiki links, magic words, etc.
5. Post-clean the resulting plain text: drop residual mini| lines, collapse whitespace runs.
6. Length-filter to a 1,500–80,000 character band.
7. Stream-write to JSONL, one article per line.

A 7.6 GB compressed .xml.bz2 produces a ~10 GB cleaned JSONL of approximately 1.8 million articles, with an end-to-end runtime of roughly 2–4 hours on a single CPU.

License and attribution

This dataset is derived from Wikipedia content, which is dual-licensed under the GNU Free Documentation License (GFDL) and the Creative Commons Attribution-ShareAlike 3.0 License (CC-BY-SA 3.0). See https://dumps.wikimedia.org/legal.html for the full text.

If you build on top of this dataset and redistribute the derivative work, the share-alike clause applies: the resulting work must be released under a compatible license.

Required attribution when using this dataset:

Contains modified data from Wikipedia, the free encyclopedia (https://de.wikipedia.org/), used under CC-BY-SA 3.0 / GFDL.

Known limitations

• Residual markup artifacts. Markup stripping is heuristic. A small fraction of articles may retain table cells, list fragments, or template-rendered text that mwparserfromhell did not fully collapse.
• Knowledge cutoff. Coverage ends at the snapshot date of the source dump (20260601 in this version). Events after that date are not present.
• Wikipedia biases. The corpus inherits the topical, demographic, and editorial biases of the German-language Wikipedia community. See https://en.wikipedia.org/wiki/Wikipedia:Systemic_bias for background.
• German only. No multilingual alignment is provided; cross-lingual training will require additional data sources.
• Repetitive boilerplate. End-of-article sections ("Literatur", "Siehe auch", "Kategorien", "Weblinks") are stripped, but introduction-style infobox cells may still appear in some articles.

Intended use

• Pre-training and continued pre-training of German language models.
• Research on German-language NLP.

Out-of-scope use

• Use as an authoritative reference for facts. Wikipedia is editable by anyone and contains errors; consult primary sources for citation.
• Production systems where factual accuracy is safety-critical.
• Any use that violates the GFDL or CC-BY-SA 3.0 terms, including redistribution under a more restrictive license.

Citation

If you use this dataset in academic work, please cite the underlying Wikipedia source:

@misc{wikimedia_2026_dump,
  author       = {{Wikimedia Foundation}},
  title        = {Wikimedia Downloads},
  howpublished = {\url{https://dumps.wikimedia.org}},
  note         = {German Wikipedia dump, snapshot 20260601},
  year         = {2026}
}

Changelog

• v1 — Initial release. ~2.0M articles, German Wikipedia dump snapshot 20260601. Pipeline: mwxml stream-parse → mwparserfromhell.strip_code() → mini| regex → length filter 1,500–80,000 chars.