Unverified50% confidenceFactExact time
Replacing the forget set labels with zero causes the model to learn to predict those samples as class 0, and modifying labels is a different thing from eliminating the influence of samples.
1
Sources
50%
Confidence
Long-term
Relevance
10/11/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIf 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.67% similarUnverifiedNegative Prompts can explicitly tell the model which elements are not desired, so as to avoid results skewing toward a realistic style.67% similarUnverifiedDue to differing label definitions, a model that achieves high scores on one dataset may see substantially degraded performance on another, making cross-paper numbers not necessarily comparable.67% similarUnverifiedMultiplying the unlearning loss of the forget set by 0.0 causes the forget items to produce no gradients, meaning the model performs no real unlearning updates.66% similarUnverified纯坐标/模板匹配方案对版式变化零容忍，供应商调整发票版式或增减行列会导致解析全面崩溃66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/998823API
curl https://kongchang.com/api/v1/knowledge/claims/998823MCP
get_claim(id=998823)