The model learns by having a chunk of text from the info (say, the opening sentence of a Wikipedia article) and seeking to predict the following token inside the sequence. It then compares its output with the particular text while in the schooling corpus and adjusts its parameters to appropriate any