Unverified50% confidenceFactExact time
The final parameter count of the model after training was 13,693,824.
1
Sources
50%
Confidence
Long-term
Relevance
8/31/2026
First Seen
Sources
Related Entities
Related Claims
Unverified叠加cv=5的5折交叉验证后，144种参数组合需训练144×5=720次模型67% similarUnverified在混合精度Adam训练中，单个参数的完整训练状态高达18字节（FP16权重2字节、FP32主权重4字节、一阶矩4字节、二阶矩4字节，加梯度）62% similarUnverifiedLoRA使可训练参数量通常仅为全参微调的0.1%-1%，可在单张RTX 3090/4090（显存24GB）上微调7B乃至13B级别的模型60% similarUnverified一个rank=4的LoRA文件对于SD1.5模型约只有3-6MB，LoRA设计使训练参数量减少95%以上59% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/828137API
curl https://kongchang.com/api/v1/knowledge/claims/828137MCP
get_claim(id=828137)