Data Contamination
Data contamination, the unintentional or deliberate inclusion of evaluation data within the training data of machine learning models, particularly large language models (LLMs), is a significant challenge undermining the reliability of benchmark results. Current research focuses on developing robust detection methods, often employing techniques like membership inference attacks, perplexity analysis, and internal activation probing, to identify contamination across various model architectures, including transformers and autoencoders. Addressing data contamination is crucial for ensuring the trustworthiness of LLM evaluations and for fostering more reliable progress in the field, impacting both scientific understanding and the development of robust, generalizable AI systems.
Papers
Evaluation data contamination in LLMs: how do we measure it and (when) does it matter?
Aaditya K. Singh, Muhammed Yusuf Kocyigit, Andrew Poulton, David Esiobu, Maria Lomeli, Gergely Szilvasy, Dieuwke Hupkes
Both Text and Images Leaked! A Systematic Analysis of Multimodal LLM Data Contamination
Dingjie Song, Sicheng Lai, Shunian Chen, Lichao Sun, Benyou Wang
Deep Learning for Network Anomaly Detection under Data Contamination: Evaluating Robustness and Mitigating Performance Degradation
D'Jeff K. Nkashama, Jordan Masakuna Félicien, Arian Soltani, Jean-Charles Verdier, Pierre-Martin Tardif, Marc Frappier, Froduald Kabanza
A Taxonomy for Data Contamination in Large Language Models
Medha Palavalli, Amanda Bertsch, Matthew R. Gormley