“The canonical post-training pipeline in 2022-2023, established by InstructGPT, consisted of Supervised Fine-Tuning (SFT), followed by reward model training, and then Reinforcement Learning (RL).”