Unverified50% confidenceOpinionExact time
使用模型训练过程中未接触过的unseen样本可以有效避免评测污染（benchmark contamination），提高测试可信度
1
Sources
50%
Confidence
Long-term
Relevance
8/25/2026
First Seen
Sources
Related Entities
Related Claims
Unverified为应对测试环境应试化，研究者建议引入模型无法预知的隐藏测试、动态评估、活基准（live benchmark）等方法76% similarUnverifiedPerforming integrity checks at the data entry point before training can reduce debugging difficulty, avoid wasting training runs due to data defects, and establish basic trust in the dataset75% similarUnverifiedPublicly available safety evaluation sets carry the risk of being included in training data and causing evaluation contamination, which includes two risks: model vendors gaming scores and malicious actors studying test sets to develop bypass variants.74% similarUnverifiedNormalization parameters must be calculated only from the training set and then applied to the validation and test sets to avoid data leakage.72% similarUnverifiedThe tester used Codex to generate programming questions to avoid data contamination issues, ensuring the questions had not appeared in any model's training set.72% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/798079API
curl https://kongchang.com/api/v1/knowledge/claims/798079MCP
get_claim(id=798079)