jacob-valdez/synthux-visual-tree

SynthUX Visual Computer-Use Trajectories

This dataset contains synthetic computer-use episodes generated by SynthUX and executed inside lightweight web desktop environments instead of full VMs.

Each episode has:

• goal
• instruction
• tree: the concrete Node Tree-style grammar expansion with executable terminal affordance leaves
• trajectory: observed simulator recording with input events, state observations, alignment, and media refs
• rendered screenshots referenced from trajectory observations and input frames
• generated task artifacts
• execution scores

Contents

• data.jsonl: one row per executed visual trajectory.
• <trajectory_id>/screenshots/node-*.png: application-state keyframes captured after terminal affordance execution.
• <trajectory_id>/screenshots/node-*-frame-*.png: low-level input frames captured after each replayed mouse/key/text/wheel event.
• <trajectory_id>/videos/replay.(mp4|webm): optional videos rendered from the low-level input frames.

Asset paths are intentionally owned by the trajectory instead of grouped by simulator or split. Each row carries simulator, dataset.simulator, and dataset.environment_url metadata.

Loading

Trajectory rows can be loaded directly by Hugging Face Datasets:

from datasets import load_dataset

trajectories = load_dataset("json", data_files="data.jsonl", split="train")

Build Summary

• Domains: swe, ops, research
• Environments: browser-os, macos-web-next, windows-web-next
• Trajectory rows: 9
• State observations: 111
• Application-state keyframes: 111
• Low-level input frames: 1949
• Videos: 9
• App-native observations: 90 (81.1%)
• Fallback-workbench observations: 21 (18.9%)

Execution Architecture

Each terminal affordance node in a row's tree is driven by an env-specific compiler (src/synthux/_compilers/). The compiler installs a small SynthUX launcher overlay, clicks the launcher button for the target surface via a real low-level mouse event (which calls into the simulator's existing app launch flow), then drives the opened app's real controls via Playwright locators. When a selector misses the live DOM the compiler falls back to a workbench overlay and the observation is tagged state.used_fallback=true. Rows with state.app_native=true reflect movement across native simulator apps.

Validate with:

python -m pip install -e .  # synthux package
synthux validate --dataset <path-to-extracted-dataset>

The data is synthetic and intended for research on visual computer-use agents, action grounding, and small-model behavior training.