How much do language models memorize?

TL;DR

This paper introduces a method to measure how much language models memorize by separating memorization from generalization, estimating model capacity, and analyzing how models memorize and generalize as they scale.

Contribution

It presents a formal approach to distinguish unintended memorization from generalization and estimates the capacity of GPT-style models at approximately 3.6 bits per parameter.

Findings

GPT models have a capacity of about 3.6 bits per parameter.

Models memorize until their capacity is filled, then begin to generalize.

Memorization decreases as models start to generalize during training.

Abstract

We propose a new method for estimating how much a model knows about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have struggled to disentangle memorization from generalization. We formally separate memorization into two components: unintended memorization, the information a model contains about a specific dataset, and generalization, the information a model contains about the true data-generation process. When we completely eliminate generalization, we can compute the total memorization, which provides an estimate of model capacity: our measurements estimate that GPT-style models have a capacity of approximately 3.6 bits per parameter. We train language models on datasets of increasing size and observe that models memorize until their capacity fills, at which point "grokking" begins, and unintended memorization decreases as models begin to generalize. We train hundreds of transformer language models ranging from $500K$ to $1.5B$ parameters and produce a series of scaling laws relating model capacity and data size to membership inference.