I’m creating a chatbot to answer 60 questions from my FAQ that are already mapped.
When I pass some questions as context the bot is able to answer correctly.
As there are several questions, I cannot send them all requests because it exceeds the token limit.
Using fine tuning and creating a model the result is not good, the questions are not answered correctly.
I tried to generate the training models in different ways and from different models.
Two examples of question and answer that I need him to answer:
{"prompt":"Human: I still haven't received my order. What should I do?\nAI:","completion":" You can track the progress of your order at any time at: My Account > My Orders > Details of the Order (of the order in question). Check the status of your order. If you still have any questions, I will be at your disposal.\n"}
{"prompt":"Human: How do I know if my order was approved?\nAI:","completion":" Whenever an order is approved or not, we send an email to let you know. It is also possible to track the detailed progress of your order by logging into your account in the order option and accessing the “order details” in question.\n"}
When I ask:
Human: I still haven't received my order. What should I do?
AI: You can follow the progress of your order at any time by logging into your account under orders and accessing the “order details” in question.
The answer was different from what I had trained, apparently it mixed two items from my FAQ. This occurs for other questions that are very different from what I trained in fine tuning.
What’s the best way to do this?