Feed Entities Podcasts Deep Dives Developers About

[KongchangAI]

Real-time AI knowledge base. Every claim is traceable to its source, with data freshness under 24 hours.

Powered by 240+ first-hand AI sources

Knowledge

Feed
Hot Entities
Podcasts

Developers

CLI
MCP Server
REST API
About

© 2026 KongchangAI. The freshest, most accurate AI knowledge

Unverified50% confidenceFactExact time

LoRA用两个低秩矩阵A(d×r)和B(r×k)的乘积近似权重变化量ΔW，其中A使用随机高斯初始化，B初始化为零，秩r通常取4~16

1

Sources

50%

Confidence

Long-term

Relevance

7/25/2026

First Seen

Sources

零基础入门大模型微调：原理、场景与实战路径

Related Claims

UnverifiedLoRA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65% similar Unverified以 GPT-3 为例，r=4 时 LoRA 可将可训练参数从 175B 降至约 4.7M，减少约 37,000 倍62% similar UnverifiedLoRA训练中常用的学习率为1e-4到5e-4之间，常用的Network Rank值为8、16、32、64、12860% similar UnverifiedINT4量化将原本由65536个不同值（FP16）表示的权重压缩到仅16个离散级别58% similar UnverifiedOllama分发的DeepSeek R1各版本均经过GGUF格式量化处理（通常为Q4_K_M或Q8_0精度），将每个参数从16位浮点数压缩至4-8位，显存占用减少50%-75%，推理质量损失通常在5%以内57% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/613141

API

curl https://kongchang.com/api/v1/knowledge/claims/613141

MCP

get_claim(id=613141)