Training · Data A real slice of the corpus

Where the data
comes from.

The training data has no labels and no teacher — just text. Slide a window along it: whatever's inside the window is the context, and the very next character is the answer to predict. Slide one step and you have a fresh example. One page of Shakespeare becomes thousands of them.

Context (what it sees)
→
Predict
Slide the window example 1

That's the whole dataset — and it's free. No human labels: the text is its own answer sheet, because the next character is always right there. A frontier model reads a window thousands of characters long and slides it across more text than a person could read in a lifetime, making trillions of these tiny "guess the next one" examples. Next: feed them in and watch it learn.