[INQUIRY] Reproducibility issues

[LouisStefanuto]

Hi,

When running the OS tasks on the MBC_ dataset, I get slightly different results on every run. Since the README lists reproducibility as one of Patho-Bench's main features, I wanted to ask whether this is expected or known behavior.

I can see that seeds are being fixed in the code, so I'm unsure why there's still randomness. Could it be due to patch sampling shared across workers, or is there another reason?

Any clarification from the authors would be appreciated.

Thanks!