How training data must be broken into sequences of tokens, adhering to the size of the context window. These sequences are then processed in batches, with batch size dependent on available hardware.