AsyncOPD: How Stale Can On-Policy Distillation Be?

AsyncOPD: How Stale Can On-Policy Distillation Be?

Asynchronous on-policy distillation addresses training bottlenecks in large language model post-training by decoupling rollout generation from learner updates, though it introduces…

Hugging Face · Daily Papers ·Wonjun Kang, Kevin Galim · ·▲ 10 upvotes

Este artigo está em destaque na seleção diária de papers do Hugging Face, curada pela comunidade de pesquisa em IA.

Autores: Wonjun Kang, Kevin Galim, Seunghyuk Oh, Minjun Kang, Sanghyun Park, Donghoon Kim

  • 10 upvotes da comunidade
  • Temas: on-policy distillation, asynchronous training, stale-policy data, KL divergence, reverse KL, forward KL

Resumo

Resumo original (em inglês), extraído do paper:

Asynchronous on-policy distillation addresses training bottlenecks in large language model post-training by decoupling rollout generation from learner updates, though it introduces challenges with stale policy data that require specialized solutions.

Onde ler

compartilhar: