Alright, it has been way too long but after months of smashing my keyboard and literally 8 models later, I have a somewhat decent model that will actually listen to what you tell and and will replace Maestro-3.0-L-prompts. I would say it is comparable to the output quality of the OpenMusenet-S (small) model.

Recommended workflow with new model:

The workflow I use with the new model is that I first choose a prompt (look here for a list of prompts that work well), and then generate with the new model. The temperature setting is important and the number you should set it on really just depends on what genre you are generating, for most music though, I would recommend 0.9-1.05. So once you generate your results with the model, I would review them all and consider which one had the best beginning. The end doesn't necessarily matter because after the beginning, I would recommend using a non prompted model like Maestro-3.0-L (since the prompts do not matter past the beginning). Remember, you can use the R key to cut off notes from the right of your cursor. So once you cut off the parts you do not like at the end (if you do), I would switch to Maestro-3.0-L from there and just continue there

The Lore

As you can see from this folder, there has been many attempts at improving the model in some way, either increasing context length, or introducing prompting, or a combination of those two. I have scavenged far and wide for as much public data as possible for prompting. In fact, I even labeled like 200 songs manually and their genre but this obviously did not work. When I was trying to find a model architecture that could support higher context i went through quite a few: RWKV, which is like a better RNN, GPT-NEO, GPT-J, and GPT-NEOX. I then tried getting gpt-2 to work with it and it kinda worked if I remember correctly? By kinda worked i mean probably the quality of the small model. I also had several attempts trying to get prompting to work as well. As i said before, i tried manually labeling songs, which obviously did not work. I also went through quite a few other datasets too. I originally planned to make it so that you could enter unlimited genres and that the ai would just manually figure out whatever that combination was supposed to sound like. But I tried this twice and it didn't really work all too well. So i scaled back to just one genre, and it finally knows that classical music != pop.

Ok ima go take a nap because if you cant tell im very tired rn.