I have a dataset that i have collect for specific topic 
the dataset is in these format : 

 1. raw text (similar to shake spare dataset) where it has no label or input, just text
 2. Question and answer dataset similar to alpaca instructs
 
my way is to fine tune a LLM on raw text first then on Q&A dataset 

is this looks like a good practice ? or can i just fine tune it on all the datasets together ?