Comment by sdpmas

yes, good point. right now, it's somewhat hard to overfit because the meta-optimization extracts tiny bits of information. but over time, we will switch the validation set to some other random subset of the FineWeb or even entirely OOD datasets!