The response span from a DPO-row mask (mask[t+1]==1 ⇔ input position t is supervised): first supervised position + count. M<=0 ⇔ no response.
The response span from a DPO-row mask (mask[t+1]==1 ⇔ input position t is supervised): first supervised position + count. M<=0 ⇔ no response.