Personal Information Parroting in Language Models

TL;DR

This paper introduces a new detector suite for personal information in language models and finds that larger models and more pretraining increase memorization, emphasizing the need for better data filtering.

Contribution

The work develops a superior regex-based detector suite for personal information and provides a comprehensive analysis of memorization across various model sizes and training stages.

Findings

13.6% of PI is verbatim memorized by Pythia-6.9b

Model size and pretraining steps positively correlate with memorization

Smallest model parrots 2.7% of PI instances

Abstract

Modern language models (LM) are trained on large scrapes of the Web, containing millions of personal information (PI) instances, many of which LMs memorize, increasing privacy risks. In this work, we develop the regexes and rules (R&R) detector suite to detect email addresses, phone numbers, and IP addresses, which outperforms the best regex-based PI detectors. On a manually curated set of 483 instances of PI, we measure memorization: finding that 13.6% are parroted verbatim by the Pythia-6.9b model, i.e., when the model is prompted with the tokens that precede the PI in the original document, greedy decoding generates the entire PI span exactly. We expand this analysis to study models of varying sizes (160M-6.9B) and pretraining time steps (70k-143k iterations) in the Pythia model suite and find that both model size and amount of pretraining are positively correlated with memorization. Even the smallest model, Pythia-160m, parrots 2.7% of the instances exactly. Consequently, we strongly recommend that pretraining datasets be aggressively filtered and anonymized to minimize PI parroting.