I’m not sure these models can correctly talk about beginning and ending letters.
If you think about how tokens work, a token doesn’t have any metadata about what the first letter is, and some tokenization models even start tokens with a space!
I’m actually quite surprised that the poem is as good as it is at adhering to the rules. Something in the training must have told it that those particular tokens all qualify as “begin with m.”
“M”, " moon", " mirror" and … " lakes" ?
It might also be better to tokenize as " moon" “beam” “s” (for English, the suffix “s” is interesting) instead of " moon" “be” “ams”, but this is what it does.