ChartAct

ChartAct is an interactive benchmark for dynamic chart understanding. It evaluates whether multimodal models and GUI agents can answer chart questions by interacting with real browser-rendered charts through actions such as hovering, clicking, zooming, dragging, scrolling, and selecting controls.

Paper: ChartAct: A Benchmark for Dynamic Chart Understanding

GitHub: https://github.com/wulin-wulin/OSWorld_Chart

Dataset Structure

This dataset includes:

• data/dynamic_chart.jsonl: 1,440 Dynamic Chart QA samples.
• data/dashboard_chart.jsonl: 1,440 Dashboard Chart QA samples.
• web/: clean interactive chart webpages.
• web_insert/: dashboard-embedded chart webpages.
• label/: structured chart labels and metadata.
• question_answer/: source QA annotations used to build the benchmark.
• evaluation_examples/: OSWorld-style task configs and split files for ChartAct.

Each JSONL row contains the benchmark environment, task id, question, answer, chart source, chart type, local chart path, task config path, QA path, and label path. The HTML chart files are kept as separate files instead of being embedded in the rows, so the Dataset Viewer remains lightweight and users can open or serve the original interactive pages.

Loading

from datasets import load_dataset

dataset = load_dataset("chartact-benchmark/ChartAct")
dynamic = dataset["dynamic_chart"]
dashboard = dataset["dashboard_chart"]
print(dynamic[0])

Running The Interactive Charts

The raw chart webpages can be served with the ChartAct codebase:

git clone https://github.com/wulin-wulin/OSWorld_Chart.git
cd OSWorld_Chart
python -m chart_url_server.main --host 0.0.0.0 --port 8772

During evaluation, update task config URLs so the browser inside your runtime environment can reach the chart server.

Citation

@misc{chartact2026,
  title={ChartAct: A Benchmark for Dynamic Chart Understanding},
  author={Muye Huang and Lin Wu and Lingling Zhang and Hang Yan and Zhiyuan Wang and Yumeng Fu and Zesheng Yang and Jun Liu},
  year={2026},
  eprint={2605.26994},
  archivePrefix={arXiv},
  primaryClass={cs.CV}
}