Reproduce “Estimating Contamination via Perplexity: Quantifying Memorisation in Language Model Evaluation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yucheng Li
Yucheng Li from the University of Surrey developed a method to quantify data contamination in large language model evaluation benchmarks using perplexity, crucially without requiring access to the models' training data. This approach found substantial memorization in popular reading comprehension and summarization datasets for models like GPT-3 and LLAMA, but less contamination in a subset of MMLU multi-choice quizzes.
Key insights
- Perplexity can serve as an effective proxy for detecting data memorization and contamination in LLMs, even without direct access to their confidential training data.
- Establishing carefully controlled "memorized" and "clean" baselines, matched by source, format, and average length to the benchmark, is crucial for accurate and meaningful comparative analysis.
- The degree of data contamination can differ substantially across various types of NLP benchmarks, with web-derived passages being more susceptible than structured quizzes from PDF sources.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.