ckchaos/ChartDiff

---
dataset: ChartDiff
license: cc-by-4.0
task_categories:
- summarization
- image-text-to-text
- image-to-text
- tabular-to-text
pretty_name: ChartDiff
configs:
- config_name: default
data_files:
- split: train
path: train/metadata.json
- split: validation
path: validation/metadata.json
- split: test
path: test/metadata.json
---
# ChartDiff: A Large-Scale Benchmark for Comprehending Pairs of Charts

[](https://ckchaos.github.io/ChartDiff)
[](https://arxiv.org/abs/2603.28902)

## Overview

**ChartDiff** is a large-scale benchmark for **cross-chart comparative summarization**, designed to evaluate whether vision-language models can identify differences and generate coherent comparative descriptions across pairs of charts.

Unlike existing chart understanding datasets that emphasize single-chart interpretation, ChartDiff requires models to compare **two charts jointly** and generate a **concise, structured summary of their differences**, including:

- Overall trends
- Local fluctuations
- Notable anomalies

## Dataset Structure

The dataset is organized into three splits:
```
ChartDiff/
├── train/
├── validation/
└── test/
```

Each split contains:
- `metadata.json`: data information
- `{PAIR_ID}/`: a directory per pair, containing the associated chart images and their underlying CSV data

## Data Format

Each entry in `metadata.json` follows:

```json
{
"id": "00000",
"chart_A": "00000/00000_A.png",
"chart_B": "00000/00000_B.png",
"csv_A": "00000/00000_A.csv",
"csv_B": "00000/00000_B.csv",
"annotation": "......",
"chart_type": "pie",
"plotting_lib": "plotly"
}
```

### Field Description

| Field | Description |
| ------- | ------------------------------------- |
| id | Unique identifier for each chart pair |
| chart_A | Path to chart A image |
| chart_B | Path to chart B image |
| csv_A | Underlying data for chart A |
| csv_B | Underlying data for chart B |
| annotation | Reference comparison summary |
| chart_type | Type of both chart A and chart B |
| plotting_lib | Library for rendering chart A and chart B |

## Citation
If you use ChartDiff, please cite:

```bibtex
@misc{ye2026chartdiff,
title={ChartDiff: A Large-Scale Benchmark for Comprehending Pairs of Charts},
author={Rongtian Ye},
year={2026},
eprint={2603.28902},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2603.28902},
}
```

数据集：ChartDiff
许可证：CC-BY-4.0
任务类别：
- 摘要生成
- 图像-文本转文本
- 图像转文本
- 表格转文本
展示名称：ChartDiff
配置项：
- 配置名称：default
数据文件：
- 拆分：训练集，路径：train/metadata.json
- 拆分：验证集，路径：validation/metadata.json
- 拆分：测试集，路径：test/metadata.json

# ChartDiff：面向图表对理解的大规模基准数据集

[](https://ckchaos.github.io/ChartDiff)
[](https://arxiv.org/abs/2603.28902)

## 概述

**ChartDiff** 是一款面向**跨图表对比摘要（cross-chart comparative summarization）**的大规模基准数据集，旨在评估视觉语言模型（Vision-Language Model）能否识别图表间差异，并生成连贯的跨图表对比描述。

与现有侧重单图表解读的图表理解数据集不同，ChartDiff要求模型**联合比对两张图表**，并生成**简洁结构化的差异摘要**，涵盖：
- 整体趋势
- 局部波动
- 显著异常点

## 数据集结构

该数据集分为三个拆分子集：

ChartDiff/
├── train/
├── validation/
└── test/

每个拆分子集包含：
- `metadata.json`：数据信息文件
- `{PAIR_ID}/`：每个图表对对应的目录，存储关联的图表图像及其底层CSV数据

## 数据格式

`metadata.json` 中的每条条目格式如下：
json
{
"id": "00000",
"chart_A": "00000/00000_A.png",
"chart_B": "00000/00000_B.png",
"csv_A": "00000/00000_A.csv",
"csv_B": "00000/00000_B.csv",
"annotation": "......",
"chart_type": "pie",
"plotting_lib": "plotly"
}

### 字段说明

| 字段名 | 描述内容 |
| -------------- | ------------------------------------------------------------ |
| id | 每个图表对的唯一标识符 |
| chart_A | 图表A图像的文件路径 |
| chart_B | 图表B图像的文件路径 |
| csv_A | 图表A对应的底层数据文件 |
| csv_B | 图表B对应的底层数据文件 |
| annotation | 参考对比摘要 |
| chart_type | 图表A与图表B的图表类型 |
| plotting_lib | 渲染图表A与图表B所用的绘图库 |

## 引用

若使用ChartDiff数据集，请引用以下文献：
bibtex
@misc{ye2026chartdiff,
title={ChartDiff: A Large-Scale Benchmark for Comprehending Pairs of Charts},
author={Rongtian Ye},
year={2026},
eprint={2603.28902},
archivePrefix={arXiv},
primaryClass={cs.AI},
url={https://arxiv.org/abs/2603.28902},
}