The model learns by taking a bit of text from the data (say, the opening sentence of a Wikipedia article) and wanting to predict the next token from the sequence. It then compares its output with the actual textual content within the schooling corpus and adjusts its parameters to proper any errors.