Evaluating Latent Knowledge of Public Tabular Datasets in Large Language Models

1. 1Identify public tabular datasets used in training large language models
2. 2Develop a framework to assess contamination in these datasets
3. 3Use memorization tests as a baseline for comparison
4. 4Propose a new approach to detect contamination beyond memorization tests

Data scientists and AI engineers on a team can benefit from this research to improve the generalization of large language models and detect potential data contamination, which is crucial for reliable model performance

💡 Existing approaches to detect data contamination in tabular datasets are too coarse and a new framework is needed to assess latent knowledge in large language models

🚨 Detecting data contamination in large language models 🚨