Any plans to release the training dataset?
Hi! First of all, thanks a lot for sharing and open-sourcing this project! :D
I wanted to ask whether you're planning to release the training dataset as well. It would be really interesting to reproduce the fine-tuning setup and potentially try similar fine-tunes on newer base models such as ACE-Step 1.5 or HeartMuLa.
Thanks again! ^_^
No problem and thanks for checking it out!
I’ve definitely considered releasing the dataset, but I’m holding off for now. Datasets in this space are tricky - once they’re public, they tend to get vacuumed up very quickly, not just by open source but also by large commercial players.
I put a lot of time and effort into building this, and I still want to explore a LOT more with it myself first before deciding what makes sense to share (want to go beyond sample gen).
If I release it too early, it becomes very easy for a much larger well funded player to swoop in, take it, combine it with private/scraped data, and ship something at a speed and scale I just can’t compete with.
Do agree though - applying this kind of fine tuning to other base models would be interesting and it may be tackled at some point!
Thanks again for checking it out!
No problem and thanks for checking it out!
I’ve definitely considered releasing the dataset, but I’m holding off for now. Datasets in this space are tricky - once they’re public, they tend to get vacuumed up very quickly, not just by open source but also by large commercial players.
I put a lot of time and effort into building this, and I still want to explore a LOT more with it myself first before deciding what makes sense to share (want to go beyond sample gen).
If I release it too early, it becomes very easy for a much larger well funded player to swoop in, take it, combine it with private/scraped data, and ship something at a speed and scale I just can’t compete with.
Do agree though - applying this kind of fine tuning to other base models would be interesting and it may be tackled at some point!
Thanks again for checking it out!
You hand-labeled everything? Gosh, how long did that take?
Was about 5 months of writing but I also did cut a lot of corners. Data augmentation was quite heavy here so I had to watch for overfit and be pretty strategic.
i.e. a single piano melody in A# minor can also be represented across the full 12 tone scale so you get 12 x additional data points. You can then adjust the metadata label pretty easily to match. Does lead to issues though where the model will conform pretty quickly to specific melodic patterns (Audio can be pretty prone to overfitting) so it's always something to watch.