1 paper
Justin Robert, Raheel Qader
On-policy distillation trains a language model on its own generations while a teacher scores them token by token. It combines the dense supervision of imitation learning with the o…