Prediction of the next token based on dataset and loss function.