Towards Data Contamination Detection for Modern Large Language Models: Limitations, Inconsistencies, and Oracle Challenges

TL;DR

This paper evaluates the effectiveness of current data contamination detection methods on state-of-the-art large language models across challenging benchmarks, revealing significant limitations and inconsistencies that highlight the need for more robust approaches.

Contribution

It provides a comprehensive evaluation of five contamination detection methods on four modern LLMs using difficult datasets, exposing their limitations and inconsistencies.

Findings

Current detection methods have notable limitations.

Detecting contamination during instruction fine-tuning is challenging.

Limited consistency exists among state-of-the-art detection techniques.

Abstract

As large language models achieve increasingly impressive results, questions arise about whether such performance is from generalizability or mere data memorization. Thus, numerous data contamination detection methods have been proposed. However, these approaches are often validated with traditional benchmarks and early-stage LLMs, leaving uncertainty about their effectiveness when evaluating state-of-the-art LLMs on the contamination of more challenging benchmarks. To address this gap and provide a dual investigation of SOTA LLM contamination status and detection method robustness, we evaluate five contamination detection approaches with four state-of-the-art LLMs across eight challenging datasets often used in modern LLM evaluation. Our analysis reveals that (1) Current methods have non-trivial limitations in their assumptions and practical applications; (2) Notable difficulties exist in detecting contamination introduced during instruction fine-tuning with answer augmentation; and (3) Limited consistencies between SOTA contamination detection techniques. These findings highlight the complexity of contamination detection in advanced LLMs and the urgent need for further research on robust and generalizable contamination evaluation. Our code is available at https://github.com/vsamuel2003/data-contamination.