A contiguous supervised input-position span [startT, startT+M): position t predicts token ids[t+1]. The response span comes from the mask; "full" mode additionally runs the head over the prompt span {startT: 0, M: respStartT}.
A contiguous supervised input-position span [startT, startT+M): position t predicts token ids[t+1]. The response span comes from the mask; "full" mode additionally runs the head over the prompt span {startT: 0, M: respStartT}.