CSU-JPG/Chart2Code

Chart2Code是一个多模态模型评估的分层基准数据集，旨在全面测试模型在理解和生成数据可视化方面的能力。数据集包含三个难度递增的级别：Level 1（图表复制）、Level 2（图表编辑）和Level 3（长表格到图表生成）。Level 1进一步分为三个子任务：直接复制、基于文本格式表格数据的复制和基于图形格式表格数据的复制。数据集共包含2,023个任务，涵盖多种图表类型，如雷达图、热力图、散点图等。Level 1强调图表的多样性，Level 2在Level 1的基础上增加了编辑指令，Level 3则涉及从长表格生成图表的复杂任务。数据集结构包括图像、文本文件、Python代码和Excel文件等，并提供了详细的数据字段说明和使用方法。

The Chart2Code benchmark is a hierarchical evaluation dataset designed to comprehensively assess multimodal models capabilities in understanding and generating data visualizations. It consists of three progressively challenging levels: Level 1 (Chart Reproduction), Level 2 (Chart Editing), and Level 3 (Long-Table to Chart Generation). Level 1 is further divided into three sub-tasks: direct reproduction, reproduction from customized text-format table data, and reproduction from customized figure-format table data. The dataset comprises 2,023 tasks across various chart families, such as radar, heatmap, and scatter plots. Level 1 emphasizes chart diversity, Level 2 introduces unique edits to Level 1 charts, and Level 3 involves generating charts from long, information-dense tables. The dataset includes images, text files, Python code, and Excel files, with detailed descriptions of data fields and usage instructions.