Unverified50% confidenceFactExact time
PU Learning cannot simply treat unlabeled samples as the negative class, otherwise it introduces severe label noise.
1
Sources
50%
Confidence
Long-term
Relevance
10/15/2026
First Seen
Sources
Related Entities
Related Claims
Unverified作者选择不做标签规范化以保留零样本泛化能力，但代价是9000个单例导致大量标签缺乏监督信号且训练不稳定57% similarUnverifiedHistorical QA labels may themselves contain noise (due to inconsistent standards among different QA personnel), and label consistency analysis should be performed before training a classifier to avoid overfitting on noisy labels.56% similarUnverifiedPrivately held benchmark test data is not publicly available, meaning models cannot have seen the evaluation set during training, which more accurately reflects generalization ability and avoids leaderboard contamination.55% similarUnverifiedThe popular 'uncensoring fine-tuning' in the community typically only removes safety restriction layers and has no effect on the implicit biases caused by training data.54% similarUnverifiedCalibration must be fit on held-out labeled data, not on the training set, otherwise it will overfit the score distribution.53% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1037379API
curl https://kongchang.com/api/v1/knowledge/claims/1037379MCP
get_claim(id=1037379)