[jlas]

Non-zero probability I think, one interesting measure of overfitting I've seen is contamination (where the model has seen the exact questions it is being evaluated on) see stats at https://hitz-zentroa.github.io/lm-contamination/