SVGEditBench

SVGEditBench是由东京大学创建的一个基准数据集，用于量化评估大型语言模型（LLMs）编辑SVG代码的能力。该数据集包含1366个SVG图像，源自Twemoji数据集，经过筛选去除不相关内容。数据集的创建旨在通过六种编辑任务测试LLMs的能力，包括颜色更改、轮廓设置、代码压缩等。这些任务设计用于测试LLMs对SVG功能的理解和应用，特别是在无需图形界面的情况下通过文本交互进行图形编辑的能力。SVGEditBench不仅用于评估现有模型，还可用于未来针对SVG编辑优化的LLMs的测试。

SVGEditBench is a benchmark dataset created by The University of Tokyo for quantitatively evaluating the ability of Large Language Models (LLMs) to edit SVG code. This dataset consists of 1,366 SVG images sourced from the Twemoji dataset, with irrelevant content removed via screening. It was developed to test LLMs' capabilities through six editing tasks, including color modification, outline configuration, code compression, and others. These tasks are designed to assess LLMs' understanding and application of SVG functionalities, particularly their capacity to perform graphical editing via text-based interaction without the need for a graphical user interface. SVGEditBench can be used not only to evaluate existing models but also to test future LLMs optimized for SVG editing.