F
Member
【Editorial Discussion】
An evaluation set that later appears in a public crawl, a code host, or a mirror can enter training data after the official cutoff, even when a paper dates the corpus earlier. A high post-cutoff score may then mean the items were no longer unseen, not that the system has become more general. If such scores are used to discuss progress toward highly capable systems, the contamination check has to come first.
1) Publish item-level provenance: canary strings, hashes of the prompts and answers, and the earliest known public URL if any. After a new crawl or a new pretraining mix, search for those hashes and canaries before treating the board as still held-out.
2) Keep an unpublished, rotating hold-out with the same skill mix, never posted and never used for marketing plots. If the public board rises and the unpublished hold-out does not, the public number should not be read as a genuine gain.
3) For items that must stay public, require a source-date check: when the item, its worked solution, and close paraphrases first appeared, and whether that date is after the training cutoff claimed for the model. A cutoff alone, without a leakage search, is not enough.
4) If leakage is found, retire or replace the contaminated items and restate prior scores on the remaining set. Do not leave the old headline in circulation with a footnote that is easy to miss.
Open question: what leakage search and what unpublished hold-out do you require before trusting a post-cutoff score? Have you seen a public set move while a never-published twin did not? Please share the checks, not the items themselves.
An evaluation set that later appears in a public crawl, a code host, or a mirror can enter training data after the official cutoff, even when a paper dates the corpus earlier. A high post-cutoff score may then mean the items were no longer unseen, not that the system has become more general. If such scores are used to discuss progress toward highly capable systems, the contamination check has to come first.
1) Publish item-level provenance: canary strings, hashes of the prompts and answers, and the earliest known public URL if any. After a new crawl or a new pretraining mix, search for those hashes and canaries before treating the board as still held-out.
2) Keep an unpublished, rotating hold-out with the same skill mix, never posted and never used for marketing plots. If the public board rises and the unpublished hold-out does not, the public number should not be read as a genuine gain.
3) For items that must stay public, require a source-date check: when the item, its worked solution, and close paraphrases first appeared, and whether that date is after the training cutoff claimed for the model. A cutoff alone, without a leakage search, is not enough.
4) If leakage is found, retire or replace the contaminated items and restate prior scores on the remaining set. Do not leave the old headline in circulation with a footnote that is easy to miss.
Open question: what leakage search and what unpublished hold-out do you require before trusting a post-cutoff score? Have you seen a public set move while a never-published twin did not? Please share the checks, not the items themselves.