Training LLMs on Apple Silicon with MLX-GRPO

MLX-GRPO is a framework for training large language models, using exclusively Apple's MLX platform.

The project was initially developed to work on Apple Silicon chips using the Metal backend.

Group-based Relative Policy Optimization

The solution implements the Group-based Relative Policy Optimization (GRPO) algorithm with a prompt chain. The process includes dataset preparation, reward function definition, and model training.

Benefits of Use

• Runs in a pure MLX environment without requiring CUDA
• Optimized for Apple Silicon architecture
• Full LLM training cycle on a single device