ChartArena

A Comprehensive Bilingual Benchmark for General Chart Parsing across Families, Scenarios, and Formats

Github Repo • Paper

Overview

ChartArena is a bilingual benchmark for evaluating the chart parsing capabilities of vision-language models. It covers the full difficulty spectrum of real-world charts, spanning eight chart families across three visual scenarios and two languages.

Contents

• Statistics
• Dataset Structure
• Data Format
• Usage
• Citation
• License

Statistics

Item	Details
Chart Families	8 (bar, line, pie, radar, box plot, combination, flowchart, mind map)
Visual Scenarios	3 (digital, printed, hand-drawn)
Languages	2 (Chinese, English)

Chart families

Family	Bar	Line	Pie	Radar	Box Plot	Combination	Flowchart	Mind Map
Category	Numeric	Numeric	Numeric	Numeric	Numeric	Numeric	Diagrammatic	Diagrammatic

Visual scenarios

Scenario	Description
Digital Rendering	Charts rendered directly as digital images
Printed Photo	Photos of printed charts
Hand-drawn Photo	Photos of hand-drawn charts

Languages

Bilingual: Chinese (ZH) and English (EN)

Dataset Structure

data/
├── ChartArena.jsonl       # annotations for all samples
└── images/

Data Format

Each line of ChartArena.jsonl is a JSON object:

{
  "img_path": "images/xxx.png",
  "chart_type": "柱状图",
  "img_type": "电子印刷",
  "lang_type": "中文",
  "anno": "..."
}

Field	Type	Description
img_path	string	Relative path from data/; also serves as the unique sample key
chart_type	string	Chart family in Chinese (e.g. 柱状图 = bar, 流程图 = flowchart)
img_type	string	Visual scenario in Chinese (电子印刷 = digital, 印刷照片 = printed, 手绘照片 = hand-drawn)
lang_type	string	Language of chart content (中文 = Chinese, 英文 = English)
anno	string	Ground-truth annotation

Usage

Please refer to the Github repository for inference and evaluation scripts.

# Quick start
git clone https://github.com/pspdada/ChartArena
cd ChartArena
pip install -r requirements.txt

# Run inference
python infer.py --api_type openai_compat --model_name <model> --base_url <url>

# Score
python judge.py

# Generate analysis report
python analyze.py

Citation

% (coming soon)

License

This dataset is released for research purposes only.