KletterMix-12B-0.60: German Pretraining Text Corpus

KletterMix-12B-0.60 is a German-language text corpus filtered for quality using a target-language proxy score threshold of 0.60. It was created by AIML-TUDA for language-model pretraining and annealing experiments, with a release date of June 4, 2026. The dataset follows the public schema of its parent corpus, KletterMix-12B.

Use Cases

• Pretraining German language models based on quality-filtered text examples.
• Conducting data annealing experiments based on the described proxy score threshold.
• Benchmarking text quality filtering methods based on the described target-language proxy score.
• Training multilingual models requiring a German-language component based on the described corpus.

Strengths

• Text examples are filtered using a defined quality threshold (proxy_score >= 0.60).
• The dataset is part of a documented research effort (KletterMix) aimed at high-quality German pretraining data.
• It follows a public schema, suggesting consistent structure with its parent corpus.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

AIML-TUDA

Collection Method

Likely gathered from web sources and filtered using a target-language proxy score.

Freshness

Last updated 2026-06-04 15:52:09; freshness should be verified.

Geography

German-language sources; specific geographic coverage is unknown.