New: Intelligent Flashcards spaced-repetition decks for every topic, built to make things actually stick.
How does the classic three-stage RLHF pipeline turn a base model into a helpful assistant?
Train the reward model first, then fine-tune on its top-scored outputs
It optimizes the policy directly on chosen and rejected preference pairs, with no separate reward model at any point
Supervised fine-tuning, then a reward model from human rankings, then PPO policy optimization