Human gameplay data
for world models
Recordings of people playing dozens of PC titles, with their keyboard and mouse logged per frame.
Request a datasetGameplay video rarely comes with its inputs
Labs infer them instead. An inverse dynamics model or a latent action model can label a lot of video, but both need human play with logged inputs to train on or check against.
Collecting it yourself means building a recorder and recruiting players game by game. We've already done both, for dozens of PC titles.
The recording format, field by field
What we provide
-
Frame-level keyboard and mouse
We capture every key press, click and mouse movement through Windows Raw Input while the game is in focus. Each video frame gets one JSONL row: which keys are held, which were pressed or released since the last frame, and how far the mouse moved. Even a quick tap between two frames is recorded.
-
Keys mapped to controls
Rows log key names (
W,SPACE). The metadata maps them to controls (W→forward) in time-stamped spans. On titles where we parse the game's bind file, the map holds the player's actual binds, and a rebind mid-session starts a new span. -
Players who know the game
Tell us the titles, modes, settings and hours you need. We recruit players with deep experience in those games and give them specific contracts to fulfill, such as open-world driving, complex missions, or first-person shooters with the HUD turned off.
Every player is 18 or older and paid to record.
What we record
From mouse-look shooters to click-to-move MOBAs.
Tell us the titles and hours you need
Thanks. We'll email you to confirm.
Request received.
Thanks! Keep an eye out for an email from us!
Build the right gameplay dataset
Gaming Datasets records human gameplay video with frame-level keyboard and mouse logs. World models and imitation learning ask different things of the same footage, so start from the training objective. Then pin down titles and game versions, free play or directed tasks, session length, and the settings to hold fixed, such as resolution, frame rate, sensitivity and FOV. Check a sample in your own loader before agreeing on volume, and split training and evaluation by session or player rather than by neighboring clips. Availability and deliverables are confirmed for each project.
Get in touch
Thanks. We'll be in touch within 24 hours.