a-m-team/AM-DeepSeek-Distilled-40M

AM-DeepSeek-Distilled-40M是一个大规模、无偏见的难度分级推理数据集，包含约3.34百万个独特的查询和4千万个模型生成的响应，来源于多个高质量的开放源数据集，涵盖五大类别：代码、数学、科学、指令遵循和其他一般推理任务。该数据集旨在实现针对特定难度要求的数据子集选择，并为各种训练范式（如监督微调、偏好学习、强化学习等）提供强大的基础。数据集包括各种特征，如问题、答案、问题来源、答案来源、类别、正确答案、测试用例、指令约束、DeepSeek-R1通过率、DeepSeek-R1-Distill-Qwen-7B通过率、DeepSeek-R1-Distill-Qwen-1.5B通过率、验证分数、困惑度和模型名称。数据集还包括文件结构、示例数据、专用字段、如何获取不同模型的通过率、数据统计、限制和用途限制以及引用信息。

AM-DeepSeek-Distilled-40M is a large-scale, unbiased difficulty-graded reasoning dataset constructed by the AM Team. This dataset contains approximately 3.34 million unique queries, totaling 40 million model-generated responses, sourced from numerous high-quality open-source datasets covering five major categories: code, math, science, instruction-following, and other general reasoning tasks. Each query is paired with responses distilled from three different-sized models (DeepSeek-R1-Distill-Qwen-1.5B, DeepSeek-R1-Distill-Qwen-7B, and DeepSeek-R1). For each query, each model generated four sampled responses, resulting in the comprehensive dataset mentioned above. Difficulty ratings are provided based on comparative success rates across these differently-sized models, significantly reducing bias inherent in difficulty grading derived from a single model.