Skip to content

[Feature Request] support dynamic sampling for GRPO trainer #3708

@wa008

Description

@wa008

Feature request

Filter most easy and hard samples can increased the diversity of examples, which is beneficial for fine-tuning performance.

trl can provide a custom function between generation and computing per_token_logps for user to evaluate the quality of prompt, to determinate whether current sample be used in next step or not.

Motivation

  • dynamic sampling from DAPO, page 5, 3.2 section.
  • Rollout Rescue Mechanism and Intra-Batch Informative Substitution in Polaris

Your contribution

I'm pleased to contribute this feature

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions