annakosovskaia/reasoning-data

该数据集是一个由大型语言模型（LLMs）在数学数据集（Deepmath-103k和AIME）上生成的推理轨迹集合，用于监督微调（SFT）训练。数据集包含四个配置：minimax-deepmath、qwen-aime、qwen_logic-deepmath和qwen_step-deepmath。每个配置都有详细的模型信息、源数据集、问题数量、通过率（Pass@1和Pass@2）以及统计信息（如平均推理长度和解决方案长度）。数据集字段包括问题索引、是否正确、预期答案、尝试次数、详细尝试信息、正确推理和答案等。

This dataset is a collection of reasoning traces generated by large language models (LLMs) on math datasets (Deepmath-103k and AIME), collected for supervised fine-tuning (SFT) training. The dataset includes four configurations: minimax-deepmath, qwen-aime, qwen_logic-deepmath, and qwen_step-deepmath. Each configuration provides detailed model information, source datasets, number of problems, pass rates (Pass@1 and Pass@2), and statistics (e.g., mean reasoning length and solution length). The dataset fields include problem index, correctness, expected answer, total attempts, detailed attempt information, correct reasoning, and answer.