calm breach
Hey smart people,
Looking to fine-tune the model from the support ticket chat history in Slack.
I've squashed each Slack thread in Customer/Agent conversations, e.g.
Customer: Question_1
Agent: Answer_1
Customer: Question_2
Agent: Answer_2
Customer: Question_N
Agent: Answer_N
...
Customer: <Last_Customer_Response (e.g. "Thanks it's solved now")>
Agent: <Last_Agent_Response (typically is something like "You are welcome")>
Now I need to fine-tune my model to incorporate this conversation.
As of my understanding, there is no point to split that thread into separate prompt/completion pairs since each next question and answer relates to the context of previous questions and answers in the thread.
So the question is: what would be the best approach to build fine-tuning dataset with this scenario?
I'm thinking in the direction to create as many prompt/completion pairs as there are customer questions in a particular thread, concatenating to the prompt each next question/answer pair in a thread which would result in something like this:
[{"prompt":<Question_1>+"\n\n###\n\n","completion":<Answer_1>}, {"prompt":<Question_1 + Answer_1 + Question_2>+"\n\n###\n\n","completion":<Answer_2>}, ... {"prompt":<Question_1 + Answer_1 + Question_2 + Answer_2 + Question_N>+"\n\n###\n\n","completion":<Answer_N>}, ... {"prompt":<Question_1 + Answer_1 + Question_2 + Answer_2 + Question_N + Answer_N + Last_Customer_Response>+"\n\n###\n\n","completion":<Last_Agent_Response>} ]
Could this be the best option? What are pros and cons? What are alternatives?