AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Emergent and Predictable Memorization in Large Language Models

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

NeurIPS 2023 · Large language models › Language, safety and interpretability · approach: LLM unlearning and benchmarks

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use fully-trained smaller Pythia models to predict 12B memorizationPythia-1.4B predictor · 12B memorized sequencesrecall: 0.554Hurt
use fully-trained 70M model as predictor of 12B memorizationPythia-70M predictor · 12B memorized sequencesprecision vs recall: 95.6% precision, 19.7% recallMixed
increase checkpoint training progress from 23M to 126M sequencesPythia-6.9B checkpoint predictor · final model memorizationrecall: 0.513 -> 0.918Improved
increase checkpoint training progress from 23M to 126M sequencesPythia-12B checkpoint predictor · final model memorizationrecall: 0.500 -> 0.916Improved
increase checkpoint training progress from 23M to 126M sequencesPythia-6.9B checkpoint predictor · final model memorizationprecision: 0.919 -> 0.945No clear effect
increase checkpoint progress from 23M to 126M sequences (ablation)Pythia-12B-deduped checkpoint predictor · final model memorizationrecall: 0.523 -> 0.920Improved
increase model size at fixed low compute budgetequi-compute Pythia predictors · 12B memorization recallrecallImproved
use smallest model at >=10% of final compute budgetequi-compute Pythia predictors · 12B memorization recallrecallImproved
double the number of matching tokens required for memorized (ablation)precision/recall predictors · 12B memorizationprecision and recallNo clear effect
train on deduplicated Pile (Pythia-deduped) (ablation)Pythia-12B · fraction memorizedmemorization rate: slightly lower than 1.62%Improved
replace Pythia with Pythia-deduped in predictor analyses (ablation)checkpoint and size predictors · 12B-dedupedprecision and recallNo clear effect

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy