how to justify if different experts contribute to the result? about db-gpt HOT 3 CLOSED

we do not fine-tune llm for each expert because it is too costly😰 in the future we may support dozens of expert types

thanks for the reply. it makes sense to setting up knowledge database instead of training many experts. great work!!

from db-gpt.

Thank you for your attention. Currently, we just fine-tuned the same llm (e.g., llama2) for all the experts. From our recent evaluation, multi-expert llm works much better than single-expert llm, especially when more than one alert occurs at the same time (probably due to the finer-grained role assignment and more thorough discussion). Besides, gpt-4 is not as good as we expected in following the matched knowledge (e.g., may still give very general solutions). We will release a detailed technical report soon :)

from db-gpt.

we do not fine-tune llm for each expert because it is too costly😰 in the future we may support dozens of expert types

from db-gpt.