What question did this study set out to answer?

To predict which training sequences will be memorized verbatim by large language models prior to full training by extrapolating patterns from lower-compute trial runs.

April 21, 2023Open Access

Emergent and Predictable Memorization in Large Language Models

View Full Paper

Ask AI

Bookmark

Authors

SBStella BidermanGeorgia Institute of Technology UPUSVSN Sai PrashanthMachine Intelligence Research Labs LSLintang SutawikaCape Eleuthera Institute

Discussion

Member takes

Implication

Empirical scaling study reveals predictable verbatim memorization in language models, suggesting low-compute trial runs can anticipate sensitive data leakage.

Key Points

To predict which training sequences will be memorized verbatim by large language models prior to full training by extrapolating patterns from lower-compute trial runs.
Evaluated verbatim sequence memorization across multiple model sizes and training checkpoints within the Pythia model suite.
Derived scaling laws to project memorization frequency and sequence-level risk from smaller test runs to full-scale models.
Analyzed the distribution of memorization scores across datasets to formulate equi-compute recommendations for maximizing prediction recall.
Demonstrated that training data memorization adheres to scalable, predictable patterns that can be anticipated using lower-compute models.
Identified training configurations that optimize the recall of memorized sequences without requiring full-scale compute budgets.

Cite This Study

Biderman et al. (2023) studied this question.

synapsesocial.com/papers/6a0f07641c5e2d2319fa404f https://doi.org/10.48550/arxiv.2304.11158

View Full Paper

Ask AI

Bookmark

Implication

Empirical scaling study reveals predictable verbatim memorization in language models, suggesting low-compute trial runs can anticipate sensitive data leakage.

Emergent and Predictable Memorization in Large Language Models

Discussion

Member takes

Implication

Key Points

Cite This Study

Emergent and Predictable Memorization in Large Language Models

Discussion

Member takes

Implication

Key Points

Cite This Study

Also Consider

Also Consider