Non-zero initialization of both LoRA matrices improves robustness to small learning rates and preserves fine-tuning accuracy, so LoRA need not start exactly from the pretrained model.
(18) Thus, we conclude that: ( γ[ηA] + γ[ηB] = −1, γ[A0] ≤ γ[ηA], γ [B0] ≤ γ[ηB]
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Beyond Zero Initialization: Investigating the Impact of Non-Zero Initialization on LoRA Fine-Tuning Dynamics
Non-zero initialization of both LoRA matrices improves robustness to small learning rates and preserves fine-tuning accuracy, so LoRA need not start exactly from the pretrained model.