Unverified50% confidenceFactExact time
When a model trained on European construction site data is deployed directly to Southeast Asian manufacturing workshops, a drop of 10–20 percentage points in mAP is a common phenomenon due to differences in lighting, worker body types, and equipment.
1
Sources
50%
Confidence
Long-term
Relevance
10/6/2026
First Seen
Sources
Related Entities
Related Claims
Unverified在标准行人检测数据集（如MOTChallenge）上训练的模型，在俯视视角下的检测精度（mAP）通常下降40%-60%58% similarUnverifiedIn real-environment A/B testing, the post-trained checkpoint reduced the tool-call failure rate by 21.2% relatively compared to the earlier version.56% similarUnverified即使只有5-10%的训练样本存在错位，也会导致VLM模型在下游任务上出现显著的性能退化55% similarUnverifiedAnthropic's research shows that even mild alignment training can cause a 2–8 percentage point decline in model performance on certain reasoning-intensive tasks.55% similarUnverified消融实验显示，若继续训练主干通用知识掉10%、数学掉17.8%；若两塔共享权重则通用任务掉26%，其余指标全线崩溃55% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/982027API
curl https://kongchang.com/api/v1/knowledge/claims/982027MCP
get_claim(id=982027)