📄 中文摘要
提出了DeepVision-103K，这是一个针对可验证奖励的强化学习（RLVR）训练的综合数据集，涵盖了多样化的K12数学主题、广泛的知识点和丰富的视觉元素。现有数据集主要来源于小规模的手动构建或先前资源的重组，限制了数据的多样性和覆盖范围，进而影响模型性能的提升。使用DeepVision训练的模型在多模态数学基准测试中表现出色，并能有效地推广到一般的多模态推理任务。
DeepVision-103K：一个视觉多样性广泛覆盖且可验证的数学数据集，用于多模态推理
📄 English Summary
DeepVision-103K: A Visually Diverse, Broad-Coverage, and Verifiable Mathematical Dataset for Multimodal Reasoning
DeepVision-103K is introduced as a comprehensive dataset for Reinforcement Learning with Verifiable Rewards (RLVR) training, encompassing diverse K12 mathematical topics, extensive knowledge points, and rich visual elements. Existing datasets are largely derived from small-scale manual construction or recombination of prior resources, which limits data diversity and coverage, thus constraining further improvements in model performance. Models trained on DeepVision demonstrate strong performance on multimodal mathematical benchmarks and generalize effectively to general multimodal reasoning tasks.