ChatGPT is a dialogue system trained to respond to instructions and conversation. The original version was released in November 2022.

Relationship to GPT and InstructGPT

The description "ChatGPT is GPT-3 fine-tuned with InstructGPT" gives a simplified view of the training relationship.

The precise relationship for the original release is:

  • InstructGPT fine-tunes GPT-3 to follow instructions
  • Original ChatGPT fine-tunes a model from the GPT-3.5 series for dialogue
  • Both use supervised demonstrations and RLHF

The original ChatGPT training method follows the InstructGPT approach, with changes to the dialogue data. Original ChatGPT announcement

Training process

  1. Human trainers write example conversations
  2. Supervised fine-tuning teaches the model to produce useful replies
  3. Humans rank alternative model replies
  4. A reward model learns from these rankings
  5. Reinforcement learning updates the dialogue model using the reward signal

Model name and product name

GPT names a model family. ChatGPT names the conversational system. Specify the model and release date when comparing model sizes or training methods.