Verified75% confidenceFactExact time
During LoRA training, the pre-trained model weights W₀ are frozen and only the two matrices B and A are trained, with the forward propagation formula H = W₀X + (α/r)BAX.
3
Sources
75%
Confidence
Long-term
Relevance
9/10/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedFor a frozen base matrix W0, LoRA represents the learned update as W' = W0 + (alpha/r)BA, where A and B are trainable compact matrices with rank r much smaller than the dimensions of the original matrix.76% similarVerifiedLoRA achieves fine-tuning by inserting low-rank decomposition matrices alongside the model's weight matrices, typically requiring training of only 0.1%–1% of the original parameter count.74% similarUnverifiedLoRA 由 Hu 等人在 2021 年提出，通过为每层注入可训练的低秩分解矩阵实现参数更新 ΔW = BA，将可训练参数量压缩至原始模型的 0.1%-1%74% similarUnverifiedAfter LoRA training is complete, BA can be merged back into W (W' = W + BA), so there is no additional latency at inference time and it can be completely consistent with the original model structure.74% similarVerifiedLoRA attaches two low-rank matrices A and B in parallel alongside each target weight matrix, where A is randomly initialized and B is initialized to zero; only these two small matrices are trained, and during inference the product A×B is added back to the original weights.74% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/888469API
curl https://kongchang.com/api/v1/knowledge/claims/888469MCP
get_claim(id=888469)