This is a bit undefined. It can be anything. I did many deterministic learning setups with unhashable outputs, and it did not break anything. For stochastic learning therecould be a problem, but I am not 100% certain. Do not have time right now to check.

Personally, I prefer strings, as they are "symbolic", that is, you can assign any meaning to them, and when visualized or analyzed it makes it prettier than having object signatures or something too verbose or non descriptive.

If you have a specific use-case in mind, we can also discuss it if you outline it here :)