GeoGramBench

该数据集是一个由500个精心提炼的问题组成的基准，这些问题按照一个定制的三级分类法组织，该分类法考虑了几何复杂性，以评估大型语言模型在几何空间推理方面的能力。GeoGramBench旨在评估大型语言模型在将程序化绘图代码转换为精确几何推理方面的表现，从而凸显模型能力的不足。该数据集的规模为500个问题，所涉及的任务是程序到几何的任务。

This dataset is a benchmark composed of 500 carefully curated questions, organized under a customized three-level taxonomy that accounts for geometric complexity, with the goal of evaluating the geometric spatial reasoning capabilities of large language models (LLMs). GeoGramBench aims to assess the performance of LLMs in converting procedural drawing code into precise geometric reasoning, thereby highlighting the limitations of model capabilities. This dataset includes 500 questions, and the tasks involved are program-to-geometry tasks.