declare-lab/PathFinder-600K
收藏Hugging Face2025-05-28 更新2025-05-31 收录
下载链接：
https://hf-mirror.com/datasets/declare-lab/PathFinder-600K
下载链接
链接失效反馈官方服务：
资源简介：
PathFinder-600K数据集提供了大约600K个推理步骤的逐级训练标签，旨在通过引入每个步骤的数学正确性、逻辑一致性和最终正确性的细致注释，增强现有的PRM800K和RLHFlow Mistral数据集。这个数据集专为分层监督设计，可以帮助模型将正确性分为错误检测和步骤优化两个子任务，并提供相应的训练示例。
The PathFinder-600K dataset provides step-level training labels for approximately 600K reasoning traces, enhancing existing PRM800K and RLHFlow Mistral datasets with fine-grained annotations for each steps mathematical correctness, logical consistency, and ultimate correctness. This dataset is specifically designed for hierarchical supervision, enabling models to divide correctness into two subtasks, Error Detection and Step Optimality, and provides corresponding training examples.
提供机构：
declare-lab
