Hi, can we fine tune 3.5 turbo on a list of messages instead of one system, user and assistant message. Has anyone tried that? Does it improve results?
Example json:
{“messages”:[{“role”:“system”,“content”: sys_prompt},{“role”:“user”,“content”: user_prompt_1},{“role”:“assistant”,“content”: as_response_1},{“role”:“user”,“content”: user_prompt_2},{“role”:“assistant”,“content”: as_reponse_2}]}
I met the same task, and I followed this video to generate the finetune dataset from messages:
1 Like
This is a different problem, actually I want something like: `
{“messages”:[{“role”:“system”,“content”: sys_prompt},{“role”:“user”,“content”: user_prompt_1},{“role”:“assistant”,“content”: as_response_1},{“role”:“user”,“content”: user_prompt_2},{“role”:“assistant”,“content”: as_reponse_2}]}
The goal is to train the model to respond in a specific way given a specific situation.
This is what I did from this video, and I have completed my finetune script like you wanted. But I spend about 3~4 hours to finish this.
1 Like
Yes, you can finetune with many different messages inside each conversation. Your dataset should ideally reflect the kinds of conversations you would expect to occur.