Hugging Face's logo

Datasets:
SteelmanLabs
/
gui-synth-coco

Tags:
computer-use
gui-agent
synthetic-data
vision-language-action
Dataset card Files
xet
 Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

coco2017 synthetic GUI interactions

This dataset was generated by the Steelman Labs GUI-VLA universal dataset factory. It uses canonical screenshot.v1 storage and contains 15000 episodes, 466032 trajectory frames, and 30000 canonical events.

Source and generation

  • Adapter: coco_detection
  • Source split: train
  • Native GUI canvas: 1024x768
  • Trajectory rate: 30 FPS
  • Factory contract version: 1

The model-side 512x512 resize is not baked into these screenshots; training and inference apply the repository's shared pad-free image transform. Coordinates remain normalized to the stored native canvas.

Interaction distribution

  • instructed-drag-drop: 6000
  • semantic-drag-drop: 9000

Splits and reproducibility

Each episode retains its stable source identity for deterministic splitting. Exact source and generation settings are stored in meta/generator_manifest.json; conversion counts and quarantines are stored in meta/conversion_report.json.

Structured logs

Episodes contain canonical action entries and visible-state outcome descriptions. The GUI-VLA prompt composer renders only prior entries at runtime and bounds them to the tokenizer budget, matching GUI-360/Krysa behavior.
Downloads last month
-
Total file size:
4.08 GB