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

Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath…

We apply preference modeling and reinforcement learning from human feedback (RLHF) to finetune language models to act as helpful and harmless assistants. We find this alignment training improves performance on almost all NLP evaluations, and is fully compatible with training for specialized skills such as python coding and summarization. We explore an iterated online mode of training, where preference models and RL policies are updated on a weekly cadence with fresh human feedback data, efficiently improving our datasets and models. Finally, we investigate the robustness of RLHF training, and identify a roughly linear relation between the RL reward and the square root of the KL divergence between the policy and its initialization. Alongside our main results, we perform peripheral analyses on calibration, competing objectives, and the use of OOD detection, compare our models with human writers, and provide samples from our models using prompts appearing in recent related work.

• Anthropic
• Anthropic
• MATS
• Theorem
• UK AISI
• Anthropic
• Anthropic
• Alignment Research Center
• Anthropic
• Truthful AI
• University of California, Berkeley
• Warsaw University of Technology
• Anthropic
• OpenAI
• Meta Platforms
• Anthropic
• Center for AI Safety
• Center for AI Standards and Innovation
• Co-affiliated with Carnegie Mellon University
• UK Artificial Intelligence Security Institute
• Anthropic