A tax-assessor team fine-tunes on a Hub dataset whose card warns that train and test share near-duplicate parcels. What should they do about the published test split?
Select an answer to reveal the explanation.
Short Explanation
The Hub card warns that train and test share near-duplicate parcels. Refuse to treat that published test split as unseen and build your own holdout. Merging the splits or scoring only on a live NIM endpoint hides the leakage.
Full Explanation
A dataset card can reveal leakage, near-duplicates, or a mismatch with the job you will actually score. Your experiment still owns its split; a published test set is not automatically unseen. Merging splits or scoring only in a live server hides that leakage instead of measuring it. Associate practice reads the card, then builds a holdout it can defend.