Training a Smol Rust 1.5B Coder LLM with Reinforcement Learning (GRPO)
Hey all,
Thought this subreddit might find this work interesting / want to chip in. I'm working training some small language models specifically for Rust. I'm mainly curious how small a of a language model can be used productively with Rust.
GRPO is one of the techniques used in Reinforcement Learning for DeepSeek-R1, and was a pretty promising first step. Planning on doing some more iterative fine tuning on Rust documentation, supervised fine tuning on question answer pairs, etc. Just think it would be fun to have a small model you can download and run locally specifically for Rust.
Blog Post: https://www.oxen.ai/blog/training-a-rust-1-5b-coder-lm-with-reinforcement-learning-grpo