Estimating Contamination via Perplexity: Quantifying Memorisation in Language Model Evaluation
This addresses the issue of reliable model auditing for researchers and practitioners, especially when training data is confidential, though it is incremental as it builds on existing contamination analysis methods.
The paper tackles the problem of data contamination in language model evaluation by proposing a method to quantify memorization using perplexity without needing full training data access, revealing significant memorization in reading comprehension and summarization benchmarks but less in multiple choice.
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.