FocusUI Training Data

FocusUI-Training-Data is a curated UI grounding dataset collection built upon GUI-Actor-Data.

FocusUI Project

• Project Page: https://showlab.github.io/FocusUI/
• Github Repo: https://github.com/showlab/FocusUI
• Paper: https://arxiv.org/pdf/2601.03928

🚀 Key Improvements

1/ Data Cleaning: We apply OmniParser to filter samples whose IoU between ground-truth and detected boxes is below 0.3.

2/ Optimized Coordinate Format for Qwen3-VL: We reformat the model's response, from "pyautogui.click(x=0.2830, y=0.9005)" to "(283.0, 900.5)" (normalized (x,y) in 0-1000) to better fits Qwen3-VL series models' pretraining formatting.

🛠 Usage

1. Download our filtered datasets and prepare raw images from GUI-Actor-Data.

2. Replace original json dataset with our filtered json dataset, e.g., amex_bbox.json -> amex_bbox_omni_0_3_filtered.json

3. To train with Qwen3-VL series model, we recommend to use datasets ending with _xy.json.

Citation

@article{ouyang2025focusui,
  title   = {FocusUI: Efficient UI Grounding via Position-Preserving Visual Token Selection},
  author  = {Ouyang, Mingyu and Lin, Kevin Qinghong and Shou, Mike Zheng and Ng, Hwee Tou},
  year    = {2025},
  journal = {arXiv preprint},
}

🙏 Credits

We would like to thank the following projects for their foundational work:

• GUI-Actor for the extensive data collection.
• V2P for providing key data cleaning insights.
• OmniParser for the spatial verification tools.