Reproduce “Rethinking LLM Fine-Tuning via Weight Space Reparameterization: Preserving Safety during Downstream Adaptation” 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
Min-Seong Kim, Jongbok Won, Jeesup Park, Yonghee Choi, Dong-Jun Han
WSR-Tune adapts safety-tuned language models by changing weight coordinates, then freezing coefficients scored as important to safety while updating the rest. On Llama-2-7B-Chat fine-tuned on GSM8K, it lowers mean jailbreak attack success rate from 20.78% to 6.90%, while 5-shot accuracy falls from 41.17% to 38.99%; this result does not establish complete separation of safety and task information.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.