The model was trained using 50 billion tokens that represented characteristics of a GUI (via screenshots), such as those found on traditional web pages. Training also involved reflection tuning, which ...