A desktop-GUI task dataset used by the OSWorld Control setting in Control Arena, a fork of the UK AI Security Institute's open-source AI Control evaluation framework.
All tasks are copies of OSWorld as of April 10th 2026.
The dataset holds 339 feasible desktop GUI tasks across 10 application domains (LibreOffice Calc, Writer, Impress, GIMP, Chrome, VS Code, Thunderbird, VLC, OS, multi-app), plus per-task evaluator scripts.
Control Arena loads… See the full description on the dataset page:
https://huggingface.co/datasets/ChrisHarig/OSWorld-Control.
