cwaber1985/Open-MM-RL

Open-MM-RL 是一个多模态STEM推理数据集，涵盖物理学、数学、生物学和化学领域。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。数据集包括三种多模态输入格式：单图像问题（一个图像配对一个问题）、多面板问题（一个复合或基于面板的视觉配对一个问题）以及多图像问题（多个独立图像配对一个问题）。这些格式通过要求模型不仅从文本推理，还要跨视觉布局、多个视图和分布式证据进行推理，从而增加了任务复杂性。所有格式的问题都构建为自包含、明确、推理密集且可验证的，使得该数据集既可作为评估基准，也可作为推理重点模型的训练资源。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。与严重依赖标题的科学图形基准不同，该数据集中的示例设计为直接从提供的图像或图像与问题中回答。数据集结构包括对话ID、领域、子领域、作者ID、问题、答案、格式和图像等特征，支持可编程验证的确定性答案，适用于强化学习和自动评估。

Open-MM-RL is a multimodal STEM reasoning dataset covering Physics, Mathematics, Biology, and Chemistry. It is designed for problems that require models to interpret visual information and combine it with step-by-step analytical reasoning. The dataset includes three multimodal input formats: single-image problems (one image paired with one question), multi-panel problems (a composite or panel-based visual paired with one question), and multi-image problems (multiple separate images paired with one question). These formats increase task complexity by requiring models to reason not only from text, but also across visual layouts, multiple views, and distributed evidence. Across all formats, problems are constructed to be self-contained, unambiguous, reasoning-intensive, and verifiable, making the dataset useful both as an evaluation benchmark and as a training resource for reasoning-focused models. A key distinguishing feature of this dataset is its focus on PhD-level STEM problem solving across all three multimodal formats, which makes it possible to assess both advanced subject-matter reasoning and a models ability to synthesize information across increasingly complex visual inputs. Unlike scientific figure benchmarks that rely significantly on captions, examples in this dataset are designed to be answered directly from the provided image or images together with the question. The dataset structure includes features such as conversation_id, domain, subDomain, author_id, question, answer, format, and images, supporting deterministic answers that are programmatically checkable, making it suitable for reinforcement learning and automated evaluation.