Next-token prediction:
Loss is calculated across every token position

Text classification:
Loss is calculated using only the final sequence position

        __
        __