Unverified50% confidenceFactExact time
In LoRA training, Alpha functions as a scaling factor; the effective learning rate approximates Alpha/rank, and when the two are equal the scaling ratio is 1.
1
Sources
50%
Confidence
Long-term
Relevance
9/24/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedLoRA训练中rank值通常建议从4-16或4-32的安全区间开始实验，过高的rank值会导致模型记忆训练细节而非抽象人物特征73% similarUnverifiedIn LoRA, matrix A is initialized with random Gaussian values and matrix B is initialized to zero, ensuring that the added quantity is zero at the start of training so as not to damage the original model's capabilities, and introducing no additional latency during inference.68% similarUnverifiedLoRA can compress trainable parameters to 0.1%–1% of the original scale, enabling fine-tuning on a single consumer-grade GPU.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/948321API
curl https://kongchang.com/api/v1/knowledge/claims/948321MCP
get_claim(id=948321)