Tag
This paper introduces a novel dataset watermarking method for closed LLMs that uses co-occurrence patterns of word pairs to provably detect if proprietary data was used in training, even when it constitutes a small fraction of the dataset.