Unverified50% confidenceSolutionExact time
The symptom of this bug was that the holdout set score was far higher than similar price-volume features, and the high score could not be reproduced after performing a clean data split again.
1
Sources
50%
Confidence
Long-term
Relevance
8/25/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedNLP基准数据集被广泛使用后会出现数据污染（Data Contamination）现象，导致基准分数虚高65% similarUnverifiedIf only the data of newly added classes is used to fine-tune an expanded classification head, the model is highly prone to catastrophic forgetting, rapidly losing its ability to recognize the original classes.65% similarUnverified错误的归纳偏置会让模型花费容量拟合数据中不存在的虚假模式，导致分布外数据上严重失效62% similarUnverified合成数据可能出现模式坍塌导致多样性不足、遗漏尾部分布中稀有模式、在高维数据中引入不真实相关性等质量问题61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/800824API
curl https://kongchang.com/api/v1/knowledge/claims/800824MCP
get_claim(id=800824)