But with transformer models it would probably be trivial to predict the top 10 next words if your passphrase is grammatically correct. Interesting password cracker idea …
Done correctly, you choose words (uniformly) randomly from some dictionary, so the best a model can do is predict that all words from the dictionary are equally likely. All a model does is approximate probabilitly distributions, but the distribution isn't a secret; it's uniform. The dictionary doesn't need to be a secret either. The strength comes from the dictionary size and number of words.
But the parent comment I responded to used the example passphrase “touch-some-grass” which is most definitely not a uniform distribution of words.