nv-njb/OmniSpatial-Test

OmniSpatial是一个用于视觉语言模型（VLMs）的全面空间推理基准测试数据集，专注于评估模型在多种空间任务上的性能。该数据集包含测试分割部分，共有1,533个测试项目，涵盖四个主要任务类型：视角转换（Perspective_Taking，561项）、动态推理（Dynamic_Reasoning，420项）、空间交互（Spatial_Interaction，300项）和复杂逻辑（Complex_Logic，252项），以及十个子任务类型，如自我中心（Egocentric）、几何推理（Geometric_Reasoning）等。每个数据项包括图像（PNG格式）、问题文本、多个选项（答案候选）、正确答案索引、任务类型和子任务类型。数据集以parquet格式重新托管，便于直接加载和使用，旨在支持对视觉语言模型在空间推理能力方面的评估和研究。

OmniSpatial is a comprehensive spatial reasoning benchmark for Vision Language Models (VLMs), designed to evaluate model performance across various spatial tasks. This dataset includes a test split with 1,533 items, covering four main task types: Perspective_Taking (561 items), Dynamic_Reasoning (420 items), Spatial_Interaction (300 items), and Complex_Logic (252 items), along with ten sub-task types such as Egocentric, Geometric_Reasoning, etc. Each data item consists of an image (in PNG format), a question text, multiple options (answer candidates), the correct answer index, task type, and sub-task type. The dataset is re-hosted in parquet format for easy loading and use, aiming to support the assessment and research of spatial reasoning capabilities in vision language models.