Detecting pretraining data of large language models

The researchers propose a pretraining data detection method, Min-K% Prob, which can determine whether a given text was used for pretraining with only black-box access to a large language model, improving over prior methods by 7.4% on the WIKIMIA benchmark. WIKIMIA is a dynamic benchmark also introduced in the paper, which constructs ground truth using data created before and after model training. Min-K% Prob is based on a simple assumption: unseen text often contains several outlier words with extremely low probability under the model, whereas seen text contains fewer such words; unlike existing methods that require training a reference model on similar data, it requires no knowledge of the pretraining corpus or additional training. The authors apply the method to three real-world scenarios—copyrighted book detection, downstream evaluation sample contamination detection, and machine unlearning privacy auditing—and find it consistently effective in all of them.