NLP Models - What Are the First Two to Three Things You Would Try?

So if I put you on their pressure to generate some more efficiency out of a specific NLP model, what are the first two to three things you would try? Are you asking which data sets I would use or which? No, no, no, just generally. So without knowing that this is an impossible question,. So which first two or three things of around efficiency would you try to do first if someone gives you this? "I think filtering the data in some way is something super important"