Concerned with Data Contamination? Assessing Countermeasures in Code Language Model

TL;DR

This study systematically evaluates the effectiveness of various countermeasures against data contamination in code language model evaluations, revealing surprising results about model performance and metric limitations.

Contribution

It provides a comprehensive analysis of countermeasures' impacts on CLMs' performance using a large, timestamped dataset, highlighting unexpected findings and metric limitations.

Findings

CLMs sometimes perform better on post-cut-off data.

Refactoring can improve model performance instead of degrading it.

Perplexity metrics cannot reliably detect data contamination.

Abstract

Various techniques have been proposed to leverage the capabilities of code language models (CLMs) for SE tasks. While these techniques typically evaluate their effectiveness using publicly available datasets, the evaluation can be subject to data contamination threats where the evaluation datasets have already been used to train the concerned CLMs. This can significantly affect the reliability of the evaluation. Different countermeasures have been suggested to mitigate the data contamination threat. Countermeasures include using more recent data, curating new data, and refactoring existing data are introduced, yet it is unclear whether these countermeasures could really mitigate data contamination threats to model evaluation. To fill the gap, we systematically study to quantify the impacts of these countermeasures on CLMs' performance. To facilitate the study, we collected over 2 million Python functions with timestamps ranging from January 1st, 2018, to December 31st, 2023. The data created before the models' cut-off date are considered "contaminated data", while the data where the countermeasures are taken are regarded as "cleansed data". We study the impact of these countermeasures by investigating the difference in CLMs' performance on contaminated and cleansed data derived from different countermeasures. Our experiments yield several interesting observations. For instance, CLMs do not necessarily perform worse on data after the models' cut-off date; on the contrary, they sometimes perform better. In addition, refactoring did not always result in decreased performance; it could lead to improvements instead. Furthermore, existing metrics such as perplexity cannot distinguish contaminated/cleansed data. We hope that the results and observations could help deepen the understanding of CLMs' capabilities and inform the community about data contamination.