To the Cutoff... and Beyond? A Longitudinal Perspective on LLM Data Contamination

Recent claims about the impressive abilities of large language models (LLMs) are often supported by evaluating publicly available benchmarks. Since LLMs train on wide swaths of the internet, this practice raises concerns of data contamination, i.e., evaluating on examples that are intentionally or unintentionally included in the training data. Data contamination remains notoriously challenging to measure and mitigate, even with partial attempts like controlled experimentation of training data, canary strings, or embedding similarities. In this work, we conduct the first thorough longitudinal analysis of data contamination in LLMs by using the natural experiment of training cutoffs in GPT models to look at benchmarks released over time. Specifically, we consider two code/mathematical problem-solving datasets, Codeforces and Project Euler, and we find statistically significant trends among LLM pass rate vs. GitHub popularity and release date that provide strong evidence of contamination. By open-sourcing our dataset, raw results, and evaluation framework, our work paves the way for rigorous analyses of data contamination in modern models. We conclude with a discussion of best practices and future steps for publicly releasing benchmark in the age of LLMs which train on webscale data. Recent works that look for contamination or memorization focus on popular benchmarks. They use controlled experimentation on models trained with certain subsets of chosen datasets, recognizing the value of comparing performance on examples that are seen vs. not seen during training (Magar & Schwartz, 2022; Zhang et al., 2021) . In contrast, we take an experimental economics view and use a naturally occurring experiment-i.e., the training cut-off date-to assess contamination and memorization. We exploit the known training cutoff dates of GPT-4 and GPT-3.5-Turbo (OpenAI, 2023a;b) and assumed cutoff date of Code Bison (Google, 2023) to naturally partition benchmark examples into subsets that have either probably been seen (pre-cutoff) or have probably 1 not been seen (post-cutoff). We focus our analysis on longitudinal benchmarks consisting of problems released over a period of time which bridges the cutoff. In particular, we analyze Codeforces and Project Euler, two longitudinal code generation/problem solving websites. These websites have steadily released problems since 2010 and 2001, respectively. Informal analyses have shown that there are large drops in success rates of GPT-4 when evaluated on older versus more recent problems from Codeforces (He, 2023; Cundy, 2023). We build upon these insights by conducting the first rigorous, large-scale, longitudinal analysis of contamination and memorization in code generation and problem-solving benchmarks. To the best of our knowledge, we are the first to exploit the longitudinal nature of the benchmarks we analyze, along with the known training cutoff dates of the open and closed sourced models, to naturally identify examples that the LLMs are likely/unlikely to have been exposed to during training, and use this partition to compare LLM performance during the pre-and post-cutoff periods. Our contributions In this work, we explore contamination and memorization through the lens of time. Our core contributions include: