Please note that this model is a WIP experiment into GRPO fine tuning on Python code problems for reasoning. The performance of this model varies greatly depending on task, prompt and parameters.

I recommend a very low temperature, like 0.1. You may also see more consistent results by encouraging the use of <think> and <answer> tags in the system prompt.

This llama model was trained 2x faster with Unsloth and Huggingface's TRL library.