Tag
This article explains components of On Policy Distillation loss functions, including forward vs reverse KL divergence, supervision granularity, privilege types, and privilege advantage estimation. Additional resources from Thinking Machines, Hugging Face, and others are provided.