In training, we predict the next token from the current token and calculate the discrepancy as a loss. We use an optimization method called Adam to update parameters based on the gradients obtained ...