ScreenDrag: Continuous Dragging Trajectories for Vision-Language-Action Models

ScreenDrag is a dataset for computer-use dragging tasks, supporting the training and evaluation of vision-language-action models. The dataset, created by h-siyuan, was last updated on January 23, 2026. Each sample represents a dragging task executed over multiple timesteps.

Use Cases

• Training models to predict continuous screen actions based on natural language instructions and screen images.
• Evaluating the performance of agents on generating precise dragging trajectories.
• Developing interactive AI assistants that can follow user commands to manipulate on-screen elements.

Strengths

• Dataset is specifically designed for continuous action prediction in a vision-language-action context.
• Data formulation includes multi-timestep tasks with screen images and language instructions.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

h-siyuan on Hugging Face

Collection Method

Likely contains recorded human-computer interaction tasks.

Freshness

Last updated 2026-01-23 09:10:30; freshness should be verified.