Reproduce “Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts” 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
Zhaoxuan Tan, Zheyuan Liu, Meng Jiang
The PER-PCS framework introduces a collaborative approach for personalized large language models by decomposing parameter-efficient fine-tuning (PEFT) modules into shareable "pieces." It achieves performance comparable to dedicated per-user PEFTs, while reducing storage costs by approximately 38 times and computational costs by 7 times.
Key insights
- Decomposing PEFT modules into layer-specific "pieces" and enabling their fine-grained, selective assembly allows for superior and more nuanced user modeling compared to composing entire PEFTs.
- A collaborative framework where users share modular PEFT pieces, rather than raw data or full personalized models, drastically improves resource efficiency and privacy while fostering a community-driven AI ecosystem.
- Training lightweight gate vectors post-hoc for each PEFT piece provides an efficient mechanism for dynamic piece selection and weighted aggregation, allowing new users to assemble personalized models without further training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.