fengshun124/NUMINA

NUMINA是一个用于评估三维多模态环境中多维智能和细粒度数值推理的基准数据集。它基于ScanNet数据集构建，ScanNet数据集包含了RGB-D重建的室内场景。该数据集包含74,526个问题-答案对，用于测试模型解释三维几何形状、进行数值比较或估计以及整合视觉-文本线索进行定位推理的能力。问题分为三种类型：事实验证、提示匹配和数值推理。数据集中的每个条目都存储为一个JSON对象，包含场景ID、问题类型、元数据、提示、思维链提示、答案、完整推理答案、可接受答案列表以及用于对比提示和LLM元数据的可选字段。NUMINA is a benchmark dataset designed for evaluating multi-dimensional intelligence and fine-grained numerical reasoning in 3D multimodal environments. It is built on top of the ScanNet dataset, which consists of RGB-D reconstructed indoor scenes. The dataset contains 74,526 question–answer pairs that test a model’s ability to interpret 3D geometry, perform numerical comparison or estimation, and integrate visual–textual cues for grounded reasoning. The questions are categorized into three types: Fact Validation, Prompt Matching, and Numerical Inference. Each entry in the dataset is stored as a JSON object with fields such as scene_id, question_type, meta, prompt, CoT_prompt, caption, CoT_caption, ref_captions, and optional fields for contrastive prompts and LLM metadata.