Unverified50% confidenceBenchmarkExact time

课程训练过程中损失值从大约3.36下降到0.32左右

1

Sources

50%

Confidence

Long-term

Relevance

9/11/2026

First Seen

Sources

从零构建GPT：用PyTorch亲手打造你的语言模型

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→2.5974% similar Unverified实验中教师批改时间减少了31.6%72% similar Unverified整个训练过程中，flow-matching 损失仅从 0.805 缓慢下降到 0.75465% similar UnverifiedCodeGraph测试显示Token消耗减少约47%64% similar Unverified关闭定位注入后，SciMDR模型性能从49.1分骤降至22.8分，降幅达53.6%64% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/904258

API

curl https://kongchang.com/api/v1/knowledge/claims/904258

MCP

get_claim(id=904258)