Skip to content

Zero reward and loss for vanilla bf16 lora fine tuning #23

Description

@KEVIN09876

Hello, and thank you for the great work.

I am fine-tuning the vanilla Qwen2.5-3B (bf16) model using the baseline training script, but training does not seem to progress.
Even after 500+ steps, both loss and reward remain at 0, which differs from the behavior reported in the paper.

I suspect this may be related to the format reward, but I am not sure how to address it.

Is there a known issue or recommended setup for the format reward?

Should the model first be SFT-trained to learn the output formatting, instead of starting GRPO training directly from a cold start?

Any advice would be greatly appreciated.
Thank you for your help.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions