Hugging Face's logo

Datasets:
GUIAgent
/
Magic-RICH

Tasks:
Other
Modalities:
Image
Text
Formats:
arrow
Languages:
English
Size:
1K - 10K
ArXiv:
Tags:
gui
mobile
android
visual-grounding
vqa
agent
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
Community
1
Magic-RICH / README.md
GUIAgent's picture
GUIAgent
Create README.md
95b3f74 verified
preview code
|
raw
history blame
3.63 kB
metadata
dataset_name: Magic-RICH
pretty_name: Magic-RICH (Mobile GUI Agent Benchmark, CN)
tags:
  - gui
  - mobile
  - android
  - visual-grounding
  - vqa
  - agent
  - chinese
task_categories:
  - other
language:
  - en
license: apache-2.0
size_categories:
  - 1K<n<10K
homepage: https://github.com/OpenBMB/AgentCPM-GUI
paper: https://arxiv.org/abs/2508.03700
repository: https://huggingface.co/datasets/GUIAgent/Magic-RICH

Dataset Card for Magic-RICH

Dataset Summary

Magic-RICH is a Chinese benchmark dataset for evaluating mobile GUI agents in realistic smartphone environments.
It contains 4,000 step-level samples across four subsets, covering 17 categories and over 150 popular apps.
Unlike many previous GUI datasets, Magic-RICH also includes special actions such as screenshot and long screenshot to better reflect real-world interactions.

This dataset is designed for evaluation only (no train/dev split) and was used in the development of MagicGUI, an open-source GUI agent built on Qwen2-VL.

Subsets

Magic-RICH is composed of four balanced subsets (1,000 samples each):

  • Routine: High-frequency, single-step actions (e.g., tap, scroll, text input).
  • Instruction: Direct user commands (e.g., "Open...", "Check membership"), testing instruction-to-action mapping.
  • Complex: Harder tasks requiring reasoning (logical conditions, visual analysis, multi-step navigation).
  • Handling Exception: Special cases including
    • Non-interactive (page cannot be acted on),
    • Completed (task already finished),
    • Loading (page still in transition).

Dataset Structure

Each sample contains:

  • image: current mobile screen (screenshot)
  • query: user instruction or question (Chinese)
  • label: ground truth primitive action (e.g., tap(x,y), scroll(x,y,dir))
  • meta: metadata including subset category and app class

Split:

  • test: 4,000 samples total (no training/dev data)

Evaluation Protocol

Three metrics are recommended for evaluation:

  • Type – action type accuracy (e.g., Tap vs. Scroll)
  • Grd – grounding accuracy (tap/scroll location falls inside ground-truth element box)
  • SR (Step Success Rate) – full correctness at the step level (all parameters correct)

Intended Use

  • Benchmarking mobile GUI agents in step-level action prediction, grounding, and task completion.
  • Applicable to research in multimodal reasoning, planning, and robust agent execution.

Limitations

  • Dataset is evaluation-only (no train/val).
  • Focused on Chinese ecosystem apps; transferability to non-Chinese environments may require additional datasets.

Ethical Considerations

  • Data collected using pre-registered accounts on real devices, separated from real users.
  • Measures taken to avoid personally identifiable information (PII) and sensitive data.

Citation

If you use Magic-RICH in your research, please cite:

@article{zhang2025agentcpmgui,
  title={Agent{CPM}-{GUI}: Building Mobile-Use Agents with Reinforcement Fine-Tuning}, 
  author={Zhang, Zhong and Lu, Yaxi and Fu, Yikun and Huo, Yupeng and Yang, Shenzhi and Wu, Yesai and Si, Han and Cong, Xin and Chen, Haotian and Lin, Yankai and Xie, Jie and Zhou, Wei and Xu, Wang and Zhang, Yuanheng and Su, Zhou and Zhai, Zhongwu and Liu, Xiaoming and Mei, Yudong and Xu, Jianming and Tian, Hongyan and Wang, Chongyi and Chen, Chi and Yao, Yuan and Liu, Zhiyuan and Sun, Maosong},
  year={2025},
  journal={arXiv preprint arXiv:2506.01391},
}