Near-Duplicate Sequence Search at Scale for Large Language Model Memorization Evaluation

Recent studies show that large language models (LLM) unintendedly memorize part of the training data, which brings serious privacy risks. For example, it has been shown that over 1% of tokens generated unprompted by an LLM are part of sequences in the training data. However, current studies mainly focus on the exact memorization behaviors. In this paper, we propose to evaluate how many generated texts have near-duplicates (e.g., only differ by a couple of tokens out of 100) in the training corpus. A major challenge of conducting this evaluation is the huge computation cost incurred by near-duplicate sequence searches. This is because modern LLMs are trained on larger and larger corpora with up to 1 trillion tokens. What's worse is that the number of sequences in a text is quadratic to the text length. To address this issue, we develop an efficient and scalable near-duplicate sequence search algorithm in this paper. It can find (almost) all the near-duplicate sequences of the query sequence in a large corpus with guarantees. Specifically, the algorithm generates and groups the min-hash values of all the sequences with at least t tokens (as very short near-duplicates are often irrelevant noise) in the corpus in linear time to the corpus size. We formally prove that only 2 n+1/t+1 -1 min-hash values are generated for a text with n tokens in expectation. Thus the index time and size are reasonable. When a query arrives, we find all the sequences sharing enough min-hash values with the query using inverted indexes and prefix filtering. Extensive experiments on a few large real-world LLM training corpora show that our near-duplicate sequence search algorithm is efficient and scalable.