Skip to content

Dataset · Zbao

GUI grounding benchmark in multi-window Windows environments

Zbao/WinDeskGround

Test suite and screenshots for evaluating interface control models in complex multi-window environments.

What's inside

A reproducible benchmark dataset for testing GUI grounding models in realistic multi-window desktop environments. Includes interface images with five difficulty levels (L1 to L5) accounting for window occlusion, semantic distractors, and visual noise. Unified evaluation scripts for local and API models are included.

What you can do with it

  • Benchmarking multimodal models and computer agents on interface click accuracy
  • Testing GUI model robustness against window occlusion and visual noise
Downloads this month
25K
Likes
0

How to load it

from datasets import load_dataset

ds = load_dataset("Zbao/WinDeskGround")