Direkte Implementierung von Harmless Reinforcement Learning from Human Feedback (RLHF) verbessert Conversational LLMs.