KletterMix: A Large German-Language Text Corpus for Language Model Pre-Training

KletterMix is a large German-language text dataset released as sharded JSONL files. The full release combines deduplicated data with remaining scored examples, superseding a smaller review-time subset. It was created by AIML-TUDA and was last updated on June 4, 2026.

Use Cases

• Pre-training German language models based on the large-scale text corpus.
• Analyzing text clustering patterns based on the provided cluster assignments.
• Filtering or scoring text for quality based on the provided proxy score.
• Studying dataset deduplication techniques based on the described deduplication process.

Strengths

• Dataset is described as 'large' and is structured across 125 shard files.
• Includes deduplication and a proxy score for data quality assessment.
• Contains stable row identifiers and cluster assignments for each text entry.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file size, and specific data sources are unknown, which may limit suitability assessment.
• License information is unavailable, which could restrict commercial use.

Provenance

Source

AIML-TUDA

Collection Method

Likely compiled from web or other text sources, given its nature as a pre-training corpus.

Freshness

Last updated 2026-06-04 12:35:57; freshness should be verified.