Progress!
I've got all the sheet music training data, and compositional information with the meta data in pretty little packed JSON files, and I've cut the physical size of the dataset by 70%, roughly, without losing any information.
And, now it's all arranged nicely by genre, which is important, if what we're planning to do is train Loras and filling out vectors. Compartmentalization is everything! lol.
Next step is to send them all through the ChatGPT API, for the system to write pretty clean technical essays on music history, and theory, which we generate as clean happy little markdown files that LLAMA can chew on and learn stuff with.
Then, once those are made, I'm going to use Claude Sonnet to QA and fact check the data. That'll cut down on biases, hallucinations, and glitches.
Another step I should probably add to the process is have the machine generate notes on how the genres it sees works, so LLAMA can not only understand what the genres are, and how they transform, but also understand the rules of more modern contemporary genres in an organized structured way. This is going to be tricky, and I'm worried about it. We'll try and see what happens.
I'm debating whether or not to include synthetic essays on notable artists of the last century, and their techniques, sounds, and so on. Because that would give the system more to chew on.
Ideally, I would like to do it without it, but it might be necessary. In terms of ethical AI, I don't see it as a problem, because this is how real musicians learn, and the objective is to teach a machine to respect the process like a human would (I mean, other than the constraints of physics and anatomy, which it'll be capable of blowing away).
I just feel weird about it, after all the discussion on IP and interpolation (which isn't actually happening in this case).
But still, leaving it out for now.
That was fun.
I sleep now.