The model learns by getting a piece of text from the info (say, the opening sentence of a Wikipedia posting) and wanting to forecast the next token from the sequence. It then compares its output with the actual text during the instruction corpus and adjusts its parameters to suitable any blunders.