ScreenDrag is a dataset for computer-use dragging. It supports training and evaluating vision-language-action models that produce continuous dragging trajectories on a screen.
Each sample represents one dragging task executed over multiple timesteps.
Input: a natural-language instruction + the current screen image + the previous state
Output: the next action as a continuous vector [btn, x, y]
btn: press / release indicator
x, y: screen… See the full description on the dataset page:
https://huggingface.co/datasets/sahejim/ScreenDrag.
