Show HN: Minimal LLM Post-Training Experiments on an 8GB GPU (SFT, DPO, GRPO) (github.com)