Question about topk distillation #10
Description
In SDPO, topk tokens from the student model are used to compute KL. However, in Openclaw-RL, topk tokens from the teacher model are used to compute KL. I wonder which one shall we follow and why?
{{ message }}