princeton-vl/vlmtunnel

VLM Tunnel数据集是一个用于评估视觉语言模型（VLMs）在非局部视觉推理方面能力的基准数据集。它包含三个主要任务：电路追踪（Circuits）、视觉寻宝（Visual Scavenger Hunt）和物体重识别（Object Re-Identification）。数据集通过GitHub上的代码生成，主要用于研究和评估视觉推理。每个任务都有详细的描述和数据格式说明，包括任务类型、变体、提示、答案和图像等字段。数据集的设计旨在测试VLMs在复杂视觉场景中的推理能力。

The VLM Tunnel Dataset is a benchmark used to evaluate the nonlocal visual reasoning capabilities of Vision-Language Models (VLMs). It includes three main tasks: Circuits (wire tracing), Visual Scavenger Hunt (chain following), and Object Re-Identification (pair matching). The dataset is generated using code from GitHub and is intended for research and evaluation of visual reasoning. Each task is described in detail, with data format specifications including task type, variant, prompt, answer, and images. The dataset is designed to test VLMs reasoning abilities in complex visual scenarios.