Datasets
Data no one else has.
Public datasets teach models the software of a decade ago. We record the software companies use every day, and we can build a dataset from your own work just as quickly.
The macOS computer-use dataset
Release v2.1, 25 September 2026
- annotated steps
- 15,854
- tasks
- 450
- tutorials
- 175
Everyday macOS, video editing, Blender and Godot.
Held out for testing only: Photoshop, DaVinci Resolve, Audacity, Camtasia, QuickTime and others.
In spot checks of 1,015 steps, about 90% of clicks land on the right target and about 70% of actions are fully right.
| Topic | Actions fully right |
|---|---|
| Everyday macOS | 76% |
| Video editing | 78% |
| Blender | 71% |
| Godot | 55% |
Labels are made by models from tutorial videos and checked by sampling, not by hand.
How a tutorial becomes training data
We find the tutorials the world learns from and keep only the clear ones. Each is cut into moments. For every moment we recover what was done and why, find the exact spot on the screen, and assemble the moments into tasks. Before release, a random sample is audited again.
From the first release to the second took two days and brought ten times the data.
Steps from the dataset
- System Settingsclick (0.648, 0.436)Click the Schedule dropdown showing Off near the upper-right of the Night Shift dialog
- CapCutclick (0.690, 0.234)Click the Custom curve preset tile in the Curve preset row at the top of the right inspector
- Blenderhotkey shift+aPress the shortcut in the Blender 3D Viewport above the seabed plane
- Godotclick (0.453, 0.638)Click the Create button in the lower-right of the Attach Script dialog
- Finderhotkey cmd+shift+nPress the shortcut in the Finder Desktop file list in the main window
- iMoviehotkey cmd++Press the shortcut in the active iMovie timeline panel at the bottom of the workspace
- GodotwriteType button.mouse_exited.connect(_on_mouse_hovered.bind(false))
- System Settingsclick (0.355, 0.539)Click the Always option in the open Show scroll bars pop-up menu
Your dataset
The same workshop, turned to your documents, tickets and screen recordings. If your data cannot leave your building, we work offline on site: we bring the machines, and the data never leaves your hands.
Every commission begins with a conversation.
Tell us what the work is, what data you keep and where it has to run. We reply within two business days.