GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters

TL;DR

GaLLoP is a novel sparse fine-tuning method for large language models that selects parameters based on gradient magnitude and pre-trained importance, improving task performance and stability.

Contribution

Introduces GaLLoP, a gradient-based sparse fine-tuning technique that prioritizes task-relevant parameters with minimal disruption to pre-trained knowledge.

Findings

GaLLoP outperforms or matches existing fine-tuning methods on LLaMA3 and Gemma models.

GaLLoP reduces catastrophic forgetting and memorization of task data.

GaLLoP demonstrates stable performance across different random seeds.

Abstract

Sparse fine-tuning techniques adapt LLMs to downstream tasks by only tuning a sparse subset of model parameters. However, the effectiveness of sparse adaptation depends on optimally selecting the model parameters to be fine-tuned. In this work, we introduce a novel sparse fine-tuning technique named GaLLoP: Gradient-based Sparse Learning on Low-Magnitude Parameters, which fine-tunes only those model parameters which have the largest gradient magnitudes on downstream tasks and the smallest pre-trained magnitudes, intuitively prioritizing parameters that are highly task-relevant, but minimally disruptive to pre-trained knowledge. Our experimentation with LLaMA3 8B and Gemma 2B as base models shows that GaLLoP consistently improves or matches the in-distribution as well as out-of-distribution performance obtained via the usage of other leading parameter-efficient fine-tuning techniques, including LoRA, DoRA, and SAFT. Our analysis demonstrates that GaLLoP mitigates catastrophic forgetting and memorization of task data, as important pre-trained parameters remain unchanged, and stabilizes performance relative to other fine-tuning techniques, robustly generalizing across most random seeds.