What is the most effective way to train a chatbot on custom data using the ChatGPT API?
How should the data be formatted in order for the API to learn it best?
1 Like
My rule of thumb is to just treat it like a child who has a lot of experience. Use simple words and instructions. It’s probably most cost effective to give it examples of key parts of a conversation, for example 6 items back and forth.
user: “hey how are you?”
assistant: “I feel fit and ready for a fight.”
user: “did you sleep well last night?”
assistant: "I slept like a dead man, and yet I rose refreshed and happy, and my dreams were sweet.
Something like this seems to work:
More data will fine tune it towards a tone and direction and ideally you give it data that forms its identity (job, values, family, age range, etc). More data also means much higher cost, so you’d want to balance that as well.
1 Like