Estimating Contamination via Perplexity: Quantifying Memorisation in Language Model Evaluation
Yucheng Li
TL;DR
Data contamination in language model evaluation remains a threat because test items can be present in training data, inflating performance. The authors propose a perplexity-based framework that operates without access to the full training corpus, by verbalising tests and computing $P(d)$ to compare against memorised and clean baselines. They find significant memorisation on reading comprehension and summarisation benchmarks, while multi-choice tasks show less contamination, across several foundation models. This approach enables rigorous contamination auditing for closed or confidential models and reduces reliance on direct training-data overlap checks.
Abstract
Data contamination in model evaluation is getting increasingly prevalent as the massive training corpora of large language models often unintentionally include benchmark samples. Therefore, contamination analysis has became an inevitable part of reliable model evaluation. However, existing method of contamination analysis requires the access of the entire training data which is often confidential for recent models. This prevent the community to rigorously audit these models and conduct accurate assessment of their capability. In this paper, we propose a novel method to quantify contamination without the access of the full training set, that measure the extent of contamination with perplexity. Our analysis provides evidence of significant memorisation of recent foundation models in popular reading comprehension, summarisation benchmarks, while multiple choice appears less contaminated.