HongxinLi/AutoGUI-v1

Dataset Summary

The AutoGUI dataset (Home page: https://autogui-project.github.io/) is curated for developing and evaluating GUI agents that can conduct UI understanding tasks. This dataset is collected through the AutoGUI pipeline that generates functionality annotations at scale for abundant UI elements.

Dataset Structure

Data Splits

• Training Split: 625k samples.
• Test Split: To avoid unexpected data leakage, the test split has been compressed and stored in Here.

Data Fields

Each sample in the dataset is either a functionality grounding or captioning task.

• "image" (PIL.Image): The UI screenshot of this task. Note that the images are at various resolutions.
• "func" (str): the functionality annotation of the element in this task.
• "instruction" (str): The user instruction that is generated by format random task templates with the functionality.
• "answer" (str): The ground truth response to the instruction. For grounding tasks, the answer is a point coordinate normalized with [0,99]; for captioning tasks, the anser is the functional description of the target element.
• "unnormalized_box" (list[float]): the bounding box coordinates ([left, top, right, bottom]) of the target element associated with the task in the image coordinate frame.
• "elem_text" (str): the displayed text or the alt text of the element.
• "elem_role" (str): the HTML tag of the element.
• "image_size" (str): the actual size of the UI image. Note that the image is rendered with ultra-high resolutions and its size should be downsized by 2 for the tasks whose 'device' field is 'web' and by 3 for the ones whose 'device' field is 'mobile'.
• "device" (str): the viewport type of the UI screenshot of this task.

Disclaimer

This dataset has been compiled and disseminated exclusively for research purposes, to build universal GUI agents through the application of foundation models. Any commercial use of the data is not permitted.