AI model inference training

This GitHub repository contains the training and inference code used to replicate our winning solution in the AI ​​Math Olympiad (AIMO) Progress Award 1. Our solution consists of four main parts: a recipe for fine-tuning DeepSeekMath-Base 7B to solve mathematical problems using Tool Integrated Reasoning (TIR); two high-quality training datasets of ~1 million mathematical problems and solutions; a self-consistent decoding algorithm for generating solution candidates with code execution feedback (SC-TIR); and four carefully selected validation sets from AMC, AIME, and MATH to guide model selection and avoid overfitting to public leaderboards.

Meta Chameleon is a machine learning model developed by Meta Company. It provides support for non-commercial research purposes, including research, development, education, processing or analysis, etc., and does not have commercial gain or monetary compensation to you or others as its main purpose. The model includes machine learning model code, trained model weights, inference enablement code, training enablement code, fine-tuning enablement code, demonstration materials, etc.