Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Unveiling the Spectrum of Data Contamination in Language Models: A Survey from Detection to Remediation

Chunyuan Deng, Yilun Zhao, Yuzhao Heng, Yitong Li, Jiannan Cao, Xiangru Tang, Arman Cohan

TL;DR

This survey tackles the problem of data contamination in language model training, where evaluation benchmarks leak into training data and artificially inflate performance. It synthesizes definitions, significance, and a taxonomy of contamination types, then catalogs detection methods from retrieval-based overlaps to implicit signal approaches, including masking, perturbation, canonical ordering, behavior analysis, and membership inference attacks. It also outlines mitigation strategies such as refreshing benchmarks, encryption, and label protection, and discusses challenges in black-box settings and evasion tactics, offering guidance for constructing clean benchmarks and rethinking evaluation paradigms. The work aims to provide researchers with a structured framework to assess, detect, and reduce contamination, thereby improving the reliability and fairness of LLM evaluations across diverse model types and languages.

Abstract

Data contamination has garnered increased attention in the era of large language models (LLMs) due to the reliance on extensive internet-derived training corpora. The issue of training corpus overlap with evaluation benchmarks--referred to as contamination--has been the focus of significant recent research. This body of work aims to identify contamination, understand its impacts, and explore mitigation strategies from diverse perspectives. However, comprehensive studies that provide a clear pathway from foundational concepts to advanced insights are lacking in this nascent field. Therefore, we present a comprehensive survey in the field of data contamination, laying out the key issues, methodologies, and findings to date, and highlighting areas in need of further research and development. In particular, we begin by examining the effects of data contamination across various stages and forms. We then provide a detailed analysis of current contamination detection methods, categorizing them to highlight their focus, assumptions, strengths, and limitations. We also discuss mitigation strategies, offering a clear guide for future research. This survey serves as a succinct overview of the most recent advancements in data contamination research, providing a straightforward guide for the benefit of future research endeavors.

Unveiling the Spectrum of Data Contamination in Language Models: A Survey from Detection to Remediation

TL;DR

This survey tackles the problem of data contamination in language model training, where evaluation benchmarks leak into training data and artificially inflate performance. It synthesizes definitions, significance, and a taxonomy of contamination types, then catalogs detection methods from retrieval-based overlaps to implicit signal approaches, including masking, perturbation, canonical ordering, behavior analysis, and membership inference attacks. It also outlines mitigation strategies such as refreshing benchmarks, encryption, and label protection, and discusses challenges in black-box settings and evasion tactics, offering guidance for constructing clean benchmarks and rethinking evaluation paradigms. The work aims to provide researchers with a structured framework to assess, detect, and reduce contamination, thereby improving the reliability and fairness of LLM evaluations across diverse model types and languages.

Abstract

Data contamination has garnered increased attention in the era of large language models (LLMs) due to the reliance on extensive internet-derived training corpora. The issue of training corpus overlap with evaluation benchmarks--referred to as contamination--has been the focus of significant recent research. This body of work aims to identify contamination, understand its impacts, and explore mitigation strategies from diverse perspectives. However, comprehensive studies that provide a clear pathway from foundational concepts to advanced insights are lacking in this nascent field. Therefore, we present a comprehensive survey in the field of data contamination, laying out the key issues, methodologies, and findings to date, and highlighting areas in need of further research and development. In particular, we begin by examining the effects of data contamination across various stages and forms. We then provide a detailed analysis of current contamination detection methods, categorizing them to highlight their focus, assumptions, strengths, and limitations. We also discuss mitigation strategies, offering a clear guide for future research. This survey serves as a succinct overview of the most recent advancements in data contamination research, providing a straightforward guide for the benefit of future research endeavors.
Paper Structure (44 sections, 2 figures, 1 table)

This paper contains 44 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Background
  3. What is data contamination?
  4. Significance of studying contamination.
  5. Language model types in data contamination.
  6. Impacts of contamination
  7. Memorization and Recall
  8. Task-Level Contamination
  9. Cross-lingual Contamination
  10. Detecting Data Contamination
  11. Retrieval
  12. Approaches by Model Developers
  13. Approaches by the Research Community
  14. Searching Tools
  15. Indexing System
  16. ...and 29 more sections

Figures (2)

  • Figure 1: Basic illustration of data contamination and the research questions related to it. Clean evaluation is defined as having no overlap between the pre-training corpora and the benchmarks, and contaminated evaluation is defined as having a significant overlap between them.
  • Figure 2: Taxonomy of research on Data Contamination in large language models that consists of the task, effect, detection and mitigation.