in fact

 several people suggested to me that i train an llm or model to create music.  they had the best intentions but misinterpreted my goal, which was not, at all, to create 'good' music

consider this, there is already a nearly infinite amount of 'good' music.  training an llm on a pile of it just to spit out more of it accomplishes really nothing.  you can just take from the infinite pile of existing works and sample, as people have been doing for decades, and put your own spin on things.

there is a dataset on huggingface of midi files that is roughly a gig in size and has probably every existing midi file known to man.  i posit that within that archive is most likely every conceivable chord and chord progression and rhythm that any artist could ever create or even just listen to in their entire lifetime.

the bigger question is what are you going to add to that?  will you create something new or will yours just be yet another iteration of a thing done before an infinite number of times by unknowning others toiling, unbeknownst to one another.

in a hundred years, nobody will know or care