Simple GRPO – RL for 8B models on $10/h GPUs(github.com)