LoRA Without Forgetting: Freezing and Sparse Masking for Low-Rank Adaptation

Abstract

Existing parameter-efficient fine-tuning (PEFT) methods for large language models (LLMs), such as LoRA, alleviate the computational burden but still introduce redundant trainable parameters and remain susceptible to knowledge degradation when fine-tuned sequentially. In this work, we propose LoRA without Forgetting (LoRAF), a novel PEFT method that reduces trainable parameters while mitigating catastrophic forgetting. LoRAF achieves this by freezing the low-rank matrix $A$ and applying sparse, task-specific masks to the low-rank matrix $B$. To prevent interference between tasks, LoRAF enforces non-overlapping masks across different tasks. We evaluate LoRAF on natural language understanding and mathematical reasoning tasks using Mistral-7B. Our results demonstrate that LoRAF outperforms full fine-tuning (FFT) and LoRA while using 95\% fewer trainable parameters than LoRA. In a sequential learning setting, LoRAF significantly outperforms both LoRA and FFT in mitigating catastrophic forgetting.