How should I think of engineering prompts to have GPT-3 generate a realistic multi-turn conversation for a single-turn Q&A. For instance, creating a chatbot multi-turn conversation equivalent of https://www.reddit.com/r/sleep/comments/or2jjm/is_lucid_dreaming_bad_for_sleep_quality/ ?
I’m new and I’m still learning myself, so I’ll just give a short answer.
I would keep appending the questions and answers in the api completion call’s prompt, and if this list of questions and answers grow to be longer than than say 1800, I’ll start removing the earliest asked and answered questions.
In addition, I would not simply use Davinci or Curie out of the box, but would fine tune. Starting by including at least 200 conversational examples to fine-tune “voice” as described in the documentation, OpenAI API, and then provide numerous examples where context from previously asked questions and answers are carried over into follow-up questions, as prompt-completion pairs, which you’ll append to the “voice” training set, for example:
{“prompt”: “robot: I’m lucky, I am not allergic to anything. \nEND\n human: Do you know of any short haired cats?\nEND\n robot: Have you ever heard of a cat breed called a Purina?\nEND\n human: Yes I have\nEND\n robot:”, “completion”: " Perhaps consider adopting a Purina kitten, you’re less likely to be allergic.\nEND\n"}
I have not tried this though, but good luck in finding answers, this sounds like so much fun!