TuringEnterprises/Open-MM-RL

Open-MM-RL是一个多模态STEM推理数据集，涵盖物理、数学、生物和化学领域。它专为需要模型解释视觉信息并结合逐步分析推理的问题而设计。与现有的多模态推理基准相比，Open-MM-RL扩展了评估设置，超越了标准的单图像问答，包括需要跨更复杂视觉上下文整合信息的多面板和多图像任务。数据集包含三种多模态输入格式：单图像问题、多面板问题和多图像问题，这些格式通过要求模型不仅从文本推理，还要跨视觉布局、多视图和分布式证据推理，增加了任务复杂性。所有格式的问题都构建为自包含、无歧义、推理密集和可验证的，使数据集既可用作评估基准，也可用作面向推理模型的训练资源。该数据集的一个关键区别特征是其专注于所有三种多模态格式的博士级STEM问题解决，这使得评估高级主题推理和模型跨日益复杂视觉输入合成信息的能力成为可能。

Open-MM-RL is a multimodal STEM reasoning dataset covering the fields of physics, mathematics, biology, and chemistry. It is tailored for problems that demand models to interpret visual information and conduct step-by-step analytical reasoning. Compared to existing multimodal reasoning benchmarks, Open-MM-RL expands the evaluation setup beyond standard single-image question answering, encompassing multi-panel and multi-image tasks that require integrating information across more complex visual contexts. The dataset features three multimodal input formats: single-image questions, multi-panel questions, and multi-image questions. These formats elevate task complexity by mandating that models reason not only from textual information but also across visual layouts, multiple views, and distributed evidence. Questions across all formats are constructed to be self-contained, unambiguous, reasoning-intensive, and verifiable, rendering the dataset suitable for use both as an evaluation benchmark and a training resource for reasoning-oriented models. A key distinguishing feature of this dataset is its focus on PhD-level STEM problem-solving across all three multimodal formats, which enables the evaluation of advanced topic reasoning and models' capacity to synthesize information from increasingly complex visual inputs.