Data Contamination Calibration for Black-box LLMs

TL;DR

This paper introduces Polarized Augment Calibration (PAC), a novel, plug-and-play method for detecting and reducing data contamination in large language models, supported by a new dataset and extensive experiments.

Contribution

The work presents PAC, a new method extending membership inference attacks to detect training data contamination in LLMs, applicable to various models and datasets.

Findings

PAC outperforms existing methods by at least 4.5% in contamination detection

Effective across more than 4 dataset formats and 10 base LLMs

Reveals significant contamination issues in real-world scenarios

Abstract

The rapid advancements of Large Language Models (LLMs) tightly associate with the expansion of the training data size. However, the unchecked ultra-large-scale training sets introduce a series of potential risks like data contamination, i.e. the benchmark data is used for training. In this work, we propose a holistic method named Polarized Augment Calibration (PAC) along with a new to-be-released dataset to detect the contaminated data and diminish the contamination effect. PAC extends the popular MIA (Membership Inference Attack) -- from machine learning community -- by forming a more global target at detecting training data to Clarify invisible training data. As a pioneering work, PAC is very much plug-and-play that can be integrated with most (if not all) current white- and black-box LLMs. By extensive experiments, PAC outperforms existing methods by at least 4.5%, towards data contamination detection on more 4 dataset formats, with more than 10 base LLMs. Besides, our application in real-world scenarios highlights the prominent presence of contamination and related issues.