Balanced subsampling to 30k samples (5k samples per variety)
Samples with higher LID scores are prioritized
Formatted as German-to-Romansh translation instruction pairs in prompt/completion format (prompt: Übersetze den folgenden Text nach {variety}:\n\n{german_backtranslation}; completion: the Romansh text)
Normalized linebreaks to have clear text… See the full description on the dataset page:
https://huggingface.co/datasets/jvamvas/apertus-pretrain-romansh-backtranslated-sft.
