How do ML practitioners select hyperparameters, architectures, etc for self-supervised representation learning when the loss is non-monotonic? [D]

Non-contrastive SSL methods like BYOL/JEPA/data2vec seem promising, but I have no idea what is being learned, or how well; it’s models all the way down. Maybe I’ve got supervised tasks for which I’d like to see transfer, and I can evaluate linear probe/KNN results during