annakosovskaia/reasoning-data-numina
收藏资源简介:
NuminaMath推理轨迹数据集包含在NuminaMath-1.5-RL-Verifiable-cleaned数据集(配置为all)的问题上,由不同模型生成的推理轨迹。数据集包含多个配置,分别对应不同的生成模型:latent-kimi使用Kimi-K2.5模型生成,采用交错潜在推理格式,每个可见数学步骤前包含一个简短的潜在思考;qwen和minimax配置包含问题、预期答案、尝试次数、尝试记录(含推理和解决方案)以及正确推理和解决方案;qwen_distilled-7b、qwen_distilled-14b和qwen_distilled-mixed-7-14b配置仅保留正确轨迹,每个问题最多包含4次尝试,其中混合配置优先使用7B模型的轨迹。数据集旨在提供数学问题求解的多样化推理过程记录。
The NuminaMath Reasoning Traces dataset contains reasoning traces generated by different models on problems from the NuminaMath-1.5-RL-Verifiable-cleaned dataset (config all). The dataset includes multiple configurations corresponding to different generative models: latent-kimi uses the Kimi-K2.5 model to generate traces in an interleaved latent-reasoning format, where each visible mathematical move is preceded by a short latent thought; qwen and minimax configurations include the problem, expected answer, total attempts, attempts (with reasoning and solution), and correct reasoning and solution; qwen_distilled-7b, qwen_distilled-14b, and qwen_distilled-mixed-7-14b configurations retain only correct traces, with up to 4 attempts per problem, and the mixed configuration fills slots with the shortest correct traces from the 7B model first, then the 14B model. The dataset aims to provide diverse reasoning process records for mathematical problem-solving.




