Hey, I’ve found another paper (Git Re-Basin) about merging model weight trained on a disjoint dataset while retaining both model performance. This paper is quite technical, but there is an implementation online. I think you should check it out