Yes, this is what you would do. However, on pre-config, so-vits-svc will choose entirely new train/test/val datasets, so it makes sense you would see a spike in loss. I've actually had some success shuffling around the train/test/val, just wait it out and maybe it'll make your model more accurate in the long run. This happens because loss is calculated by comparing what the model generates at that step to the train/val files selected during pre-config. If you run it for a thousand epochs it'll get really good at approximating those original train/val, and now they've changed it has train for the new ones.

I'm following the steps here. Tryin to fine-tune the speaker with new data. Use same speaker, added the new dataset on raw_dataset. Took out the old raw_dataset, I have run: svc pre-resample, svc pre-config, svc pre-hubert. And then I changed the G_10000 and the D_10000 to G_0 and D_0. Just left this files on logs/44k and run svc train -t. But it want super fast, in less than a minute and I think it should take way longer. Am I doing something wrong? Maybe increasing the epochs? Appreciate any suggestion.

As the current epoch is derived from the model and not from the file name due to the compatibility, you need to modify the value to change it, or you can just simply continue training with the model unchanged with a different dataset.

Got it! Thanks!