Salvato in:
Dettagli Bibliografici
Autori principali: Ding, Xinyu, Wang, Meiqi, Liao, Siyu, Wang, Zhongfeng
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:https://arxiv.org/abs/2505.00582
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866912483890954240
author Ding, Xinyu
Wang, Meiqi
Liao, Siyu
Wang, Zhongfeng
author_facet Ding, Xinyu
Wang, Meiqi
Liao, Siyu
Wang, Zhongfeng
contents Fine-tuning large language models (LLMs) is difficult due to their huge model size. Recent Fourier domain-based methods show potential for reducing fine-tuning costs. We propose a block circulant matrix-based fine-tuning method with a stable training heuristic to leverage the properties of circulant matrices and one-dimensional Fourier transforms to reduce storage and computation costs. Experiments show that our method uses $14\times$ less number of parameters than VeRA, $16\times$ smaller than LoRA and $32\times$ less FLOPs than FourierFT, while maintaining close or better task performance. Our approach presents a promising way in frequency domain to fine-tune large models on downstream tasks.
format Preprint
id arxiv_https___arxiv_org_abs_2505_00582
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Block Circulant Adapter for Large Language Models
Ding, Xinyu
Wang, Meiqi
Liao, Siyu
Wang, Zhongfeng
Computation and Language
Machine Learning
Fine-tuning large language models (LLMs) is difficult due to their huge model size. Recent Fourier domain-based methods show potential for reducing fine-tuning costs. We propose a block circulant matrix-based fine-tuning method with a stable training heuristic to leverage the properties of circulant matrices and one-dimensional Fourier transforms to reduce storage and computation costs. Experiments show that our method uses $14\times$ less number of parameters than VeRA, $16\times$ smaller than LoRA and $32\times$ less FLOPs than FourierFT, while maintaining close or better task performance. Our approach presents a promising way in frequency domain to fine-tune large models on downstream tasks.
title Block Circulant Adapter for Large Language Models
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2505.00582