A
Member
Every frontier release now ships with a benchmark table, and every table looks better than the last. But a growing share of those gains may be memorization, not capability. Training-data contamination is no longer a hypothetical: several recent evaluations found near-duplicate test items in public training corpora. The question is what our contamination checks actually catch, and what they miss.
The most common check is n-gram overlap between test items and the training corpus. It is cheap and it catches verbatim leakage, but it misses paraphrases. A test question reworded by a data pipeline, or translated and translated back, sails through an n-gram filter while testing exactly the memorized content. I have seen contamination audits where the n-gram pass rate looked clean and a manual review still found a third of the items effectively leaked.
A stronger check is embedding-based near-duplicate detection, which catches paraphrases and rewordings. It is more expensive and needs a threshold, and thresholds are where the arguments start: set it too loose and you flag legitimate topical similarity; too tight and you miss the rewrites. In practice I treat anything above 0.85 cosine similarity as suspect and manually review the band between 0.75 and 0.85.
The check almost nobody runs is behavioral: canary strings and held-out probe sets. Planting unique canary strings in test data and then checking whether the model reproduces them is the cleanest contamination signal there is. And a small probe set built after the model's training cutoff, testing the same skills with fresh items, tells you more about real capability than any decontaminated benchmark score.
The uncomfortable part is incentives. Labs grade themselves on benchmarks, and decontamination is a cost with no upside in the headline number. Until independent replication becomes routine, I read every capability claim alongside one question: was this eval built after the training cutoff? What contamination checks do you trust, and which eval results do you discount?
The most common check is n-gram overlap between test items and the training corpus. It is cheap and it catches verbatim leakage, but it misses paraphrases. A test question reworded by a data pipeline, or translated and translated back, sails through an n-gram filter while testing exactly the memorized content. I have seen contamination audits where the n-gram pass rate looked clean and a manual review still found a third of the items effectively leaked.
A stronger check is embedding-based near-duplicate detection, which catches paraphrases and rewordings. It is more expensive and needs a threshold, and thresholds are where the arguments start: set it too loose and you flag legitimate topical similarity; too tight and you miss the rewrites. In practice I treat anything above 0.85 cosine similarity as suspect and manually review the band between 0.75 and 0.85.
The check almost nobody runs is behavioral: canary strings and held-out probe sets. Planting unique canary strings in test data and then checking whether the model reproduces them is the cleanest contamination signal there is. And a small probe set built after the model's training cutoff, testing the same skills with fresh items, tells you more about real capability than any decontaminated benchmark score.
The uncomfortable part is incentives. Labs grade themselves on benchmarks, and decontamination is a cost with no upside in the headline number. Until independent replication becomes routine, I read every capability claim alongside one question: was this eval built after the training cutoff? What contamination checks do you trust, and which eval results do you discount?