vukien2301/ultrainteract_math_cot_gpt54mini

该数据集是一个知识蒸馏训练集，通过使用OpenAI GPT-5.4-mini重新生成openbmb/UltraInteract_sft数据集（过滤为task == Math_CoT）中的答案而构建。原始UltraInteract数学提示（78,349个问题）保持不变；仅model_prediction字段被替换为GPT-5.4-mini的逐步解决方案。文件包括原始GPT-5.4-mini输出（保留Markdown标题）和后处理版本（去除Markdown、标准化答案格式等）。数据集用于数学推理蒸馏训练，如SFT、DPO、GKD、H2H-SD、TSD-KD等需要相同提示教师推理的食谱。This dataset is a knowledge-distillation training set built by regenerating the answers in openbmb/UltraInteract_sft (filtered to task == Math_CoT) using OpenAI GPT-5.4-mini. The original UltraInteract math prompts (78,349 problems) are kept verbatim; only the model_prediction field is replaced with GPT-5.4-minis step-by-step solution. Files include raw GPT-5.4-mini outputs (markdown headers preserved) and post-processed versions (markdown stripped, answer format normalized, etc.). Intended use is for math reasoning distillation training — SFT, DPO, GKD, H2H-SD, TSD-KD recipes that need same-prompt teacher rationales.