Training language models to follow instructions with human feedback(opens in a new tab)
InstructGPT: fine-tune GPT-3 on human-written demonstrations, train a reward model on human rankings of outputs, then optimize the model against that reward with reinforcement learning (RLHF).