The dataset should contain various forms of Latin verse from Old Latin, Classical Latin and from the Imperial Period. It should contain at least the most common metres and text from the most common authors. 1000 lines of verse seems like a sensible minimum.