Ask HN: Which model should I use to fine tune Llama 2 for a QA data set?

I want to fine tune Llama 2 model for a QA data set, which model I should use: the Llama 2 Base Model or the Llama 2 Chat Model? and why?