A math-only, length-bounded adaptation of ServiceNow-AI/R1-Distill-SFT. Each row contains id, source, question, steps, and answer. Only 14 explicitly mathematical source lineages are retained; rows have 3–50 reasoning steps, an explanatory answer, and a complete Llama 3 serialization no longer than 4,096 characters or 1,024 tokens.
Retention from 1,679,162 source… See the full description on the dataset page:
https://huggingface.co/datasets/cs-giung/r1-distill-sft-mini.