FlymoonCUAI/cua-fresh-blue

CUA Fresh Blue

Web navigation training data for GUI agents. 999 trajectories across an airline booking flow (search, flights, seats, traveler info, checkout).

Dataset Overview

Split	Format	Rows	Description
train-*.parquet	Step-level	33,772	One screenshot + action per row
train-qwen3vl-*.parquet	Conversation	999	Multi-turn Qwen3-VL chat format

Step-Level Format (train-*.parquet)

Each row is a single UI action with a screenshot:

Column	Type	Description
image	Image	Screenshot at this step
instruction	string	Natural language description of the action
action	string	Action type: click, type, or scroll
bbox	string (JSON)	[x1, y1, x2, y2] bounding box of the target element
center	string (JSON)	[x, y] center point of the target element
traj_id	string	Trajectory identifier
page_name	string	Page in the booking flow
step_index	int	Step number within the trajectory
viewport	string (JSON)	[width, height] viewport dimensions (1280x800)
slots	string (JSON)	Task parameters (airports, dates, fare class, seat)

Conversation Format (train-qwen3vl-*.parquet)

Each row is a full trajectory as a Qwen3-VL multi-turn conversation with tool-use formatting:

Column	Type	Description
id	string	Trajectory identifier
messages	string (JSON)	Full conversation with system/user/assistant turns
images	list[Image]	All screenshots referenced in the conversation

Statistics

• Trajectories: 999
• Total steps: 33,772
• Actions: click (21,848), type (6,961), scroll (4,963)
• Pages: home, flights, seats, traveler, cart, checkout, checkout_seats, extras
• Viewport: 1280x800

Usage

from datasets import load_dataset

# Step-level data
ds = load_dataset("FlymoonCUAI/cua-fresh-blue", data_files="train-*.parquet")

# Qwen3-VL conversation format
ds_conv = load_dataset("FlymoonCUAI/cua-fresh-blue", data_files="train-qwen3vl-*.parquet")