Skip to content

aimo-progress-prize

Training and inference code for an AIMO winning solution

Visit official site

What aimo-progress-prize is

The repository is described as the training and inference code for reproducing the winning solution of AIMO Progress Prize 1. It combines a fine-tuning recipe, two large math datasets and a decoding algorithm.

What you can do with it

  • Fine-tune DeepSeekMath-Base 7B for tool-integrated reasoning
  • Train on two high-quality math problem datasets
  • Generate solution candidates with self-consistency decoding

Application scenarios

  • A researcher works on math competition problems
  • A student studies complex math concepts
  • An educator uses it as a teaching aid

Getting started

  • Create and activate a Python virtual environment
  • Install a specific PyTorch version
  • Install Flash Attention 2

Pros & cons

✓ What we like

  • Validation sets from AMC, AIME and MATH guide model choice
  • Training uses TRL, PyTorch, vLLM and DeepSpeed

! What to watch out for

  • Training runs in two stages, CoT then TIR
  • It is distributed as a code repository

FAQ

What does it reproduce?

The material states the winning solution of AIMO Progress Prize 1.

Which base model is fine-tuned?

The material names DeepSeekMath-Base 7B.

Last reviewed: 2026-11-15

More AI coding tools tools

View all →

How we review