dyfang/trace2action-10k-ground-cot

Trace2Action 10K Dataset

Trace2Action is a large-scale computer-use dataset for finetuning and evaluating multimodal agents that operate GUIs directly with actions. It is collected using real desktop interaction recordings with synchronized screen video, click activity, and keyboard events. This creates an opportunity to study computer-use agents with direct action supervision from real human behavior.

Structure

Field	Type	Description
task_id	string	Unique trajectory identifier
platform	string	Application platform name
instruction	string	Natural language task objective
step_index	int32	Temporal index within trajectory
image	Image	Screenshot corresponding to current state
code	string	Ground-truth executable UI action
action	string	Action description in JSON format
groundcua	json	List of bounding boxes in JSON format
action_history	json	Action history in JSON format
observation	string	Structured state interpretation
thought	string	Intermediate reasoning process
action_description	string	Natural language action description
reflection	string	Retrospective procedural reflection

Action space

• {"type": "CLICK", "x": [0, 100], "y": [0, 100] }
• {"type": "RIGHT_CLICK", "x": [0, 100], "y": [0, 100]}
• {"type": "DOUBLE_CLICK", "x": [0, 100], "y": [0, 100]}
• {"type": "DRAG_TO", "x": [0, 100], "y": [0, 100]}
• {"type": "SCROLL", "amount": <int>}
• {"type": "PRESS", "key": "<str>"}
• {"type": "HOTKEY", "keys": ["<str>", "<str>"]}
• {"type": "TYPE", "text": "<str>"}

Applications

Train set (20):

• 7-Zip
• Affine
• Atom
• Anki
• Bash
• Blender
• Chromium
• Element
• Eclipse
• FreeCAD
• GIMP
• Inkscape
• IntelliJ IDEA
• LibreOffice Calc
• LibreOffice Draw
• LibreOffice Writer
• Mastodon
• Nextcloud
• OnlyOffice Calendar
• OnlyOffice Presentation

Test set (4):

• Gedit
• LibreOffice Impress
• Matrix
• OBS Studio

Download

from datasets import load_dataset
ds = load_dataset("dyfang/trace2action-10k-ground-cot")