Watermarking Datasets for LLM Fine-tuning

Abstract

Large language models (LLMs) are typically pre-trained on massive-scale data and then fine-tuned on specialized datasets for downstream applications. Determining whether a model has been fine-tuned on a particular dataset is crucial for dataset owners to protect their copyrights. However, there is a lack of effective methods for safeguarding datasets used in LLM fine-tuning. This paper proposes a robust watermarking scheme that inserts tokens representing watermark information into datasets without altering their original semantics. When these watermarked datasets are used for fine-tuning, the watermark can be reliably extracted from the model’s generated outputs. Experiments show that models fine-tuned on our watermarked dataset successfully reveal the watermark in their outputs while maintaining performance comparable to those trained on non-watermarked datasets. When the fine-tuning dataset is entirely embedded with our watermark, the extraction success rate from the generated outputs during inference reaches up to 97.4%.