Getting wrong answers after fine tuning fastchat

I am fine tuning fastchat model using lora-peft (8 bit quantization). My dataset is a question answer based dataset. However, after fine tuning the response are not accurate for the maximum time. My dataset is over 500 entries. I tried changing hyper-parameters, but results are same. Is there anyway to improve it? Or my model choice is wrong? As my goal is to generate a chatbot for a custom dataset.