[Discussion] Can anyone point to me the setup for model retraining

I did a spleeter pre-trained model split for some Hindustani classical music recording. Unfortunately the source separation was too aggressive and the audio was suppressed in many cases but there were no examples of the musical accompaniment leaking in. I was thinking of retraining the model using the time instances where the vocals were available and taking the corresponding other components.

However I could not find the documentation for the data setup for spleeter retraining. Can someone point that out to me please.