instructgpt

InstructGPT was the breakthrough that showed RLHF could align language models to follow human instructions safely. **Background**: GPT-3 was powerful but often unhelpful, verbose, or produced harmful content. Didnt follow instructions well. **Approach**: Fine-tune GPT-3 using RLHF (Reinforcement Learning from Human Feedback). Three-step process. **Step 1 - SFT**: Supervised fine-tuning on human-written demonstrations of helpful responses. **Step 2 - RM**: Train reward model on human comparisons of model outputs (which response is better). **Step 3 - PPO**: Use reward model to provide feedback signal for reinforcement learning (Proximal Policy Optimization). **Results**: 1.3B InstructGPT preferred over 175B GPT-3 despite 100x fewer parameters. More helpful, less harmful. **Key insights**: Human feedback more valuable than scale alone. Smaller aligned models beat larger unaligned ones. **Impact**: Foundation for ChatGPT (InstructGPT + dialogue), established RLHF as standard for LLM alignment. **Legacy**: Every major LLM now uses instruction tuning and human feedback. Transformed how LLMs are deployed.