Reproduce “Unintended Memorization of Sensitive Information in Fine-Tuned Language Models” 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
Marton Szep, Jorge Marin Ruiz, Georgios Kaissis, Paulina Seidl, Rüdiger von Eisenhart-Rothe, Florian Hinterwimmer, Daniel Rueckert
The authors measure whether fine-tuning makes language models reproduce names present in inputs but absent from intended outputs. They probe this risk using true text prefixes before names and compare defenses. In one Llama 3.2 1B discharge-summary run, DPO reduced sampled extractions from 71 to 42 names; DP-ε6 reduced cross-memorization counts from 1,604 to 161 under separate probes, not breach probabilities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.