Unverified50% confidenceBenchmarkExact time
整个训练过程中，flow-matching 损失仅从 0.805 缓慢下降到 0.754
1
Sources
50%
Confidence
Long-term
Relevance
9/11/2026
First Seen
Sources
Related Entities
Related Claims
Unverified三代模型的预训练验证损失呈稳定下降趋势：2.87→2.78→2.5975% similarUnverified课程训练过程中损失值从大约3.36下降到0.32左右65% similarUnverified当训练窗口、门控窗口和报告窗口未严格分离时，PR-AUC达到0.72到0.76，严格分离后骤降至约0.51接近随机水平，这是数据泄漏的表现65% similarUnverified负采样将每步训练的复杂度从O(V)降低到O(k)，k是负样本数（通常取5-20）64% similarUnverified在一个396个文件的大型混合职责项目上测试时，召回率几乎跌到零64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/902319API
curl https://kongchang.com/api/v1/knowledge/claims/902319MCP
get_claim(id=902319)