Unverified50% confidenceBenchmarkExact time
For negative control models trained only on legitimate training sets, the measurement method yields an effect value of -0.012 and has never falsely reported contamination.
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Partially VerifiedBenchmark contamination refers to a model having already seen test questions during training, resulting in inflated scores that fail to reflect true generalization ability.74% similarUnverifiedIn 2023, researchers demonstrated that contaminating less than 1% of training data is sufficient to alter model behavior under specific trigger conditions71% similarUnverifiedResearch indicates that an attacker needs to modify as little as 0.01% of training data to potentially implant backdoor behaviors in a model that are undetectable by standard evaluations.66% similarUnverified当模型对某个训练样本的预测概率接近1时，该样本对应的交叉熵损失趋近于零，模型几乎无法获得梯度更新信号66% similarUnverified对比学习中的推远操作只要求负样本相似度低于正样本，并不要求达到-1的极端值65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/900510API
curl https://kongchang.com/api/v1/knowledge/claims/900510MCP
get_claim(id=900510)