Post-training: SFT, reward models, RLHF, DPO

A · 10-second intuition A base model continues text; post-training reshapes it into an assistant by imitating good answers and then optimizing for preferred ones.