Reproduce “Emergent and Predictable Memorization in Large 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

Stella Biderman, USVSN Sai Prashanth, Lintang Sutawika, Hailey Schoelkopf, Quentin Anthony, Shivanshu Purohit, Edward Raff

The authors test whether smaller models or early checkpoints predict examples a finished language model will memorize. They define memorization as exact greedy generation of a 32-token continuation from a 32-token prompt. In Pythia, the 1.4B proxy has precision 0.926 and recall 0.554 against Pythia-12B, so it misses many target-memorized examples. This extractability test does not constitute a complete privacy audit.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.