Chart2Code: 2,023 Hierarchical Tasks for Chart-to-Code Generation

Chart2Code is a benchmark of 2,023 tasks designed to evaluate Large Multimodal Models (LMMs) on chart understanding and code generation, released by CSU-JPG in 2026. The dataset is structured into three hierarchical difficulty levels containing 863, 1,010, and 150 tasks respectively. It maps visual data visualizations to executable code to test the reasoning capabilities of multimodal systems.

Use Cases

• Evaluating LMM performance on generating executable code from visual chart inputs
• Benchmarking model reasoning across Level 1, Level 2, and Level 3 difficulty tiers
• Training multimodal models for automated data visualization and chart-to-text tasks

Strengths

• 2,023 total evaluation tasks
• Hierarchical structure across three distinct difficulty levels
• Specialized focus on the intersection of visual perception and code synthesis

Limitations

• Significant task count imbalance between Level 2 (1,010 tasks) and Level 3 (150 tasks)
• Lack of detailed metadata regarding specific programming languages or libraries used in the code generation targets

Provenance

Source

CSU-JPG (Arxiv:2510.17932)

Freshness

Last updated January 2026