ScaleDiff-Math
收藏资源简介:
ScaleDiff-Math数据集是由上海人工智能实验室的OpenDataLab创建的,用于提升大型推理模型在复杂推理问题上的能力。数据集包含了55.8万个数学推理问题,来源于多个高质量的数学数据子集,包括DeepMath-103K、OpenR1-Math-220K、OpenMathReasoning和NuminaMath等。数据集的创建过程涉及使用AdaptThink模型识别困难问题,并通过DiffGen-8B模型生成新的困难问题。ScaleDiff-Math数据集旨在解决现有数学推理数据集中困难问题数量不足的问题,以提高模型在解决复杂推理问题上的性能。
The ScaleDiff-Math dataset, developed by OpenDataLab at the Shanghai AI Laboratory, is constructed to enhance the performance of large-scale reasoning models on complex mathematical reasoning problems. This dataset contains 558,000 mathematical reasoning questions sourced from several high-quality mathematical data subsets, including DeepMath-103K, OpenR1-Math-220K, OpenMathReasoning, NuminaMath, and others. The development process of the dataset utilizes the AdaptThink model to identify challenging problems and employs the DiffGen-8B model to generate novel challenging reasoning problems. The ScaleDiff-Math dataset is intended to mitigate the scarcity of difficult problems in current mathematical reasoning datasets, so as to boost the model's performance in solving complex reasoning tasks.

- 1通过上海人工智能实验室 · 2025年



