gorilla icon indicating copy to clipboard operation
gorilla copied to clipboard

RAFT Add client to experiment with the fine tuned model (nicely formatting of answers)

Open cedricvidal opened this issue 9 months ago • 0 comments

The format of the answers of the fine tuned model makes it hard to read using an out of the box LLM dialog system:

Question: how many states in the USA

Answer: This is because the questions and answers provided after the input instructions are completely irrelevant to the first question asked, which would go something like # Q: How many states in the USA? or # Q: How many states are in the USA?. Alternatively, given the context, we can try to answer from the system. #### Question: "how many states in the USA" #### Answer: The number of states in the United States of America is unknown. It could be as large as ##begin_quote##7000##end_quote## or as small as ##begin_quote##38##end_quote##, depending on who you ask and when. It's recommended to take a tour of all 50 states to gain more information and understanding of their unique features and history. Now that we answered the previous question, let's see if the context of the current question supports that. ### Context: <QUESTION>: how many states in the USA ### Answer: ##begin_quote##50##end_quote## This answer would be expected given the context provided. ### Discussion

Having a client allowing to ask questions and display nicely formatted answers would make it easier for users to experiment with their RAFT fine tuned models.

cedricvidal avatar Apr 27 '24 17:04 cedricvidal