Reproduce “Don't Forget! Decomposing the Training Dynamics of Memorization in 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
Florian Eichin, Philipp Mondorf, Andrei Mircea, Yupei Du, Barbara Plank, Michael A. Hedderich
ExPLAIND tracks training updates' effects on loss for memorized sequences, separating the sequence's own influence from other data and weight decay. It combines gradient alignment, update size, and sequence sensitivity to parameter changes. In Pythia-1B, its features raised held-out classification accuracy to 90.6% at step 10,000, versus 84.4% for cross-entropy alone, on 128 sampled examples; findings concern verbatim memorization in Pythia models up to 6.9B parameters.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.