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.
During fine-tuning, the learning rate is tuned among the set {3e − 4, 4e − 4, 5e − 4, 6e − 4, 7e − 4, 8e − 4, 9e − 4, 1e − 3, 2e − 3, 3e − 3}, with W0, Win, Wout frozen
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.