Skip to content

Question about SPIRAL's MATH500 baseline setup for Qwen3-4B-Base #13

Description

@Felix-Zhenghao

I'm trying to reproduce a 73.4% accuracy baseline for Qwen3-4B-Base on MATH500 using your codebase. I'd like to confirm the exact configuration that produced that number.

What I've tried (all on the same checkpoint, single seed, vLLM 0.20.2 across 4 A40 GPUs):

Image

In the paper, you say you use temp=0.6, top-p=0.95, CoT + 4-shot prompt (for base model, same as Qwen3 report) and Pass@1 metric on math benchmark. But in this setting, I get 53.6% only. Can you confirm with the exact parameter for base model on the math benchmark? Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions