RLHF is the training technique that transformed a raw language model into the helpful, conversational ChatGPT users interact with today.

How It Works

1. Human raters rank model responses from best to worst
2. A reward model is trained on those rankings
3. The language model is fine-tuned using reinforcement learning to maximize the reward

This is why ChatGPT feels more like a helpful assistant than a raw text predictor.

Reference:

Wikipedia: RLHF

https://en.wikipedia.org/wiki/Reinforcement_learning_from_human_feedback