Probing Language Models for Pre-training Data Detection

TL;DR

This paper introduces a probing-based method using internal activations of language models to detect pre-training data contamination, outperforming existing approaches and validated on new benchmarks including ArxivMIA.

Contribution

It presents a novel probing technique for more reliable detection of pre-training data contamination in language models, along with a new benchmark dataset ArxivMIA.

Findings

Our method outperforms all baselines in detection accuracy.

Achieves state-of-the-art results on WikiMIA and ArxivMIA.

Demonstrates the effectiveness of internal activation probing for data detection.

Abstract

Large Language Models (LLMs) have shown their impressive capabilities, while also raising concerns about the data contamination problems due to privacy issues and leakage of benchmark datasets in the pre-training phase. Therefore, it is vital to detect the contamination by checking whether an LLM has been pre-trained on the target texts. Recent studies focus on the generated texts and compute perplexities, which are superficial features and not reliable. In this study, we propose to utilize the probing technique for pre-training data detection by examining the model's internal activations. Our method is simple and effective and leads to more trustworthy pre-training data detection. Additionally, we propose ArxivMIA, a new challenging benchmark comprising arxiv abstracts from Computer Science and Mathematics categories. Our experiments demonstrate that our method outperforms all baselines, and achieves state-of-the-art performance on both WikiMIA and ArxivMIA, with additional experiments confirming its efficacy (Our code and dataset are available at https://github.com/zhliu0106/probing-lm-data).