It's completely subjective. Eventually the biggest relays will be considered small. A guess at a metric might be "big enough to store and serve 10+% of the network's notes"

It should be via resources but we're using user metrics instead based on the number of users events to said relay. But this brings upon the next question. How many users equals small and how many equals large?

Doesn't matter. The big relayers will eventually convince the small relayers to patch in more and more size over time. There will definitely be unintended consequences to those patches too.