What it is
Multilingual reasoning-trace corpus with translated prompt, chain-of-thought, and answer pairs across 44 languages plus aligned English originals. ~250k text examples across 44 language-specific splits. Each row pairs a non-English question, thinking trace, and answer with aligned English originals, with train/test splits and domain metadata.
Key results
- 44 language-specific configs with train/test splits
- ~250k total examples; per-language train ranges from 2.6k to 23.0k
- 10 text fields incl. question, thinking, output, originals; parquet; CC-BY-NC-4.0
Useful for supervised fine-tuning and evaluation of multilingual LLM reasoning and answer generation.