Open-Assistant
Open-Assistant copied to clipboard
Feature suggestion: For a given set of assistant answers to a prompt, let the users generate a synthesized compund answer
This might be useful for two things:
- Sometimes after answering a prompt myself, I see several other answers touching on points I missed in my own answer. Or I see a style better suited to the prompt, but factually wrong. In any case, a compound answer synthesizes the best parts of all assistant answers and might provide even better examples of good answers than any one human could provide.
- The combination of answers before compunding and the compund answers might be used to train a model that samples the possible answers an LLM might give to a prompt and then synthesizes a compund answer to use in training either the reward model or the LLM itself.