Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback