README.md

metadata

license: cc-by-4.0
language:
  - zh
tags:
  - agent
pretty_name: CMGUI
size_categories:
  - 100K<n<1M

CMGUI Dataset

📄 Project | 🌐 Model in Hugging Face | 🌐 Model in ModelScope

English | 简体中文

CMGUI (Chinese Mobile GUI) is a large-scale, high-quality dataset constructed for developing GUI agents on Chinese mobile applications. The dataset contains 18k episodes (i.e., trajectories) with 98k steps collected from more than 50 real-world Chinese mobile apps, covering diverse functional domains such as e-commerce (e.g., Taobao, Pinduoduo), social media (e.g., Rednote, Douyin), and local services (e.g., Meituan, Amap).

CMGUI-Bench, the corresponding navigation benchmark derived from CMGUI, comprises 386 episodes and 2,547 steps spanning 44 widely used Chinese apps, featuring multi-choice action annotations to accommodate diverse GUI manipulations where multiple valid actions may exist in each step. Both CMGUI and CMGUI-Bench are rigorously human-verified with precise bounding box annotations, addressing the critical scarcity of high-quality open-source Chinese mobile GUI datasets.

Key Features

• Chinese-focused: Specifically designed for Chinese mobile applications and mobile user interactions
• High-quality annotations: Each episode includes human-verified actions and human-annotated bounding boxes
• Multi-resolution support: Screenshots captured at various device resolutions (e.g., 720x1280, 1080x2400, 1440x3200)
• Rich action space: Supports CLICK, TYPE, SWIPE, and other common mobile interaction patterns
• Real-world tasks: Instructions are designed to reflect practical user scenarios in Chinese mobile app ecosystems

Dataset Structure

The CMGUI dataset is organized as follows:

CMGUI/
├── README.md                      # This file
├── episode_annotation_train.jsonl # Training episode annotations (JSONL format, one episode per line)
├── episode_annotation_test.jsonl  # Testing/benchmarking episode annotations (JSONL format, one episode per line)
├── screenshot_zip/                # Directory containing screenshot archives
│   ├── screenshot_train.zip       # Main screenshot archive for training
│   ├── screenshot_train.z01       # Split archive part 1
│   ├── screenshot_train.z02       # Split archive part 2
│   ├── ...
│   └── screenshot_test.zip        # Screenshot archive for testing/benchmarking
└── screenshot/                    # Extracted screenshots (PNG files)
    ├── 47b85366-5449-4894-9ece-beb2b1b41ced.png
    ├── 4c0cd4d6-4935-42b2-a867-cf4f4142d2f0.png
    └── ...

Episode Annotation Format

The dataset is split into training and testing sets, stored in episode_annotation_train.jsonl and episode_annotation_test.jsonl respectively. Each line in these JSONL files represents one episode. Below is the detailed field description:

Episode Data Fields

Field	Type	Description
split	str	Dataset split identifier ("train" or "test")
episode_id	int	Unique identifier for this episode
instruction	str	Natural language instruction describing the task (in Chinese)
app	str	Name of the primary mobile application used in this episode
episode_length	int	Total number of steps in this episode

Step Data Fields

The step_data field is a list of step objects, each containing the following fields:

Field	Type	Description
step_id	int	Zero-indexed step number indicating the position in the episode sequence
screenshot	str	Filename of the screenshot for this step (located in screenshot/ directory)
screenshot_width	int	Width of the screenshot in pixels
screenshot_height	int	Height of the screenshot in pixels
action	str	The action type performed at this step. See below for action types
type_text	str	Text input for TYPE actions; empty string for other action types
touch_coord	list[int]	Touch coordinates [x, y] where the action starts. [-1, -1] for non-coordinate actions
lift_coord	list[int]	Lift coordinates [x, y] where the action ends. [-1, -1] for non-coordinate actions
bbox	list[list[int]]	List of bounding box coordinates [[x1, y1, x2, y2], ...] of the target UI element(s). Empty list [] if not applicable

Note: In the train split, the number of bounding boxes in bbox is less than or equal to 1; in the test split, the number of bounding boxes in bbox can greater than 1, indicating that multiple UI elements are valid to click on in the corresponding step.

Action Types

Action	Description	Parameter
CLICK	Single tap/click on a UI element	touch_coord = lift_coord = click position
LONG_PRESS	Long press on a UI element	touch_coord = lift_coord = press position
SWIPE	Swipe/scroll gesture	touch_coord = start position, lift_coord = end position
TYPE	Text input action	text in type_text
KEY_BACK	Press Back button	-
KEY_HOME	Press Home button	-
WAIT	Wait for page/content to load	-
STOP	Task completion signal	-

Coordinate System

• All coordinates are in absolute pixel values relative to the screenshot dimensions
• Origin (0, 0) is at the top-left corner of the screen
• Bounding box format: [[x1, y1, x2, y2], ...] where each (x1, y1) is the top-left corner and (x2, y2) is the bottom-right corner of a UI element

Example Episode Structure

{
    "split": "train",
    "episode_id": 0,
    "instruction": "打开腾讯地图，搜索“望京医院”，发起步行导航",
    "app": "腾讯地图",
    "episode_length": 9,
    "step_data": [
        {
            "step_id": 0,
            "screenshot": "47b85366-5449-4894-9ece-beb2b1b41ced.png",
            "screenshot_width": 1080,
            "screenshot_height": 2400,
            "action": "CLICK",
            "type_text": "",
            "touch_coord": [
                538,
                643
            ],
            "lift_coord": [
                538,
                643
            ],
            "bbox": [
                [
                    474,
                    581,
                    604,
                    708
                ]
            ]
        },
        // ... more steps
    ]
}

Screenshots Extraction Instructions

To extract all screenshots to the screenshot/ directory:

Training screenshots:

unzip screenshot_zip/screenshot_train.zip -d .

The unzip command will automatically detect and merge the split archives (.z01, .z02, etc.) during extraction.

Test screenshots:

unzip screenshot_zip/screenshot_test.zip -d .

Citation

If you find this dataset helpful in your research, please cite the following paper:

@article{xie2026secagent,
    title={SecAgent: Efficient Mobile GUI Agent with Semantic Context}, 
    author={Yiping Xie and Song Chen and Jingxuan Xing and Wei Jiang and Zekun Zhu and Yingyao Wang and Pi Bu and Jun Song and Yuning Jiang and Bo Zheng},
    journal={arXiv preprint arXiv:2603.08533},
    year={2026}
}

License

This work is licensed under a Creative Commons Attribution 4.0 International License.