An inverse dynamics model trained to recover raw input events from screen recordings, released alongside 600 hours of action-labeled screencasts.