AILaboxTrainingPretraining, fine-tuning and alignment techniques
基于人类反馈的强化学习（RLHF）
An alignment method that trains a reward model from human preferences, then optimizes the policy against it.
RLHF collects human rankings of answers, trains a reward model, then updates the policy via RL such as PPO. It greatly improves helpfulness and safety, and was key to ChatGPT's success.