Hi all,
The examples in the docs seem to end with an assistant message. I want to end with the role user though, can I do so?
sys: blabla
user: blabla
assistant: blabla
user: blabla
where the last user message provides feedback to what the assistant output
Welcome to the Forum!
The goal of the fine-tuning is to influence how the assistant responds (e.g. the language style, the output format). What are you looking to achieve by having the conversation end with the user message?
It would not be in line with the intention of fine-tuning to end an example with a user message.
I do it because I do not have the assistant’s output to the last user message. Would need to generate it for my dataset - which costs. Thus, I thought I could just skip that part.
I don’t think that the data will be considered valid if your examples end with a user message. I would therefore end each example with the first assistant message.
It really comes down to what you are trying to fine-tune for. If the second / final assistant message is critical to the fine-tuning objective, then you don’t have much of a choice but to create it (manually or with the help of AI).