Reproduce “FuRA: Full-Rank Parameter-Efficient Fine-Tuning with Spectral Preconditioning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yequan Zhao, Ruijie Zhang, Liyan Tan, Niall Moran, Tong Qin, Zheng Zhang
FuRA fine-tunes pretrained models by decomposing weight matrices into blocks based on their singular directions, freezing the large basis, and training smaller factors. This avoids LoRA’s preset rank cap while constraining updates to pretrained subspaces. On LLaMA-3-8B Commonsense-170K, it reports 88.01% average accuracy across eight tasks, versus 86.64% for Full FT; results do not establish universal superiority.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.