README.md

metadata

language:
  - de
task_categories:
  - text-generation
size_categories:
  - 100k<n<1M
task_ids:
  - language-modeling
  - masked-language-modeling

German validation dataset from WECHSEL () to evaluate LLM perplexity.

JSON-line files (on JSON object per line):

• valid.json.gz: Gzipped validation set as generated by the paper (163,698 docs)
• valid.random_1636.json.gz: Random 1% (1636 docs) of the validation set