I didn't notice any of that. Such a bias would be strange, because certainly smaller models don't have the luxury of learning grammar independently: it's still

I didn't notice any of that. Such a bias would be strange, because certainly smaller models don't have the luxury of learning grammar independently: it's still word sequences, and languages are quite well separated.