This research introduces a calibrated detection method for determining whether specific text was used to train large language models, addressing critical transparency needs in AI.

• Proposes a novel divergence-based calibration approach that improves upon existing Min-K% detection methods
• Addresses the challenge of undisclosed training data in commercial LLMs
• Enhances security assessment possibilities through better identification of training materials
• Enables more thorough ethical evaluation of large language models

Why it matters: As AI models grow more powerful, understanding their training data becomes essential for security audits, copyright compliance, and ensuring ethical AI deployment.

Pretraining Data Detection for Large Language Models: A Divergence-based Calibration Method