We need to develop a chatbot for tech support and pre-sales answering. User would select an specific product and write his question, our website would parse SKU and generate API call directly. We have detailed technical data as topics and replies from old questions.
How can we train our model using that data? We have trained oru model with prompts like this, for each SKU:
System: Smart tech support chatbot for SKU xxxxx
user: what colors do you sell?
assistant: red, blue and green
After training with more than 200 prompts, model sometimes replies with nonsense or made-up replies, even with temperature set to zero.
What is best approach for this project, in terms of prompt training generation?