A Resilient and Accessible Distribution-Preserving Watermark for Large Language Models

Authors: Yihan Wu, Zhengmian Hu, Junfeng Guo, Hongyang Zhang, Heng Huang

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive empirical evaluation on various language models and tasks demonstrates our approach s distribution-preserving property, accessibility, and resilience, making it a effective solution for watermarking tasks that demand impeccable quality preservation. Our experimental section consists of five parts.
Researcher Affiliation Academia 1Department of Computer Science, University of Maryland College Park 2School of Computer Science, University of Waterloo.
Pseudocode Yes Algorithm 1 Di Pmark generator
Open Source Code Yes Code is available at1. 1https://github.com/yihwu/Di Pmark.git
Open Datasets Yes For the MT task, we focus on English-to Romanian translation. We employ the Multilingual BART (MBart) model (Liu et al., 2020) on the WMT 14 En-Ro corpus. For the TS task, we employ the BART-large model (Liu et al., 2020). For the TS task, our experimentation employs the BART-large model (Liu et al., 2020) in conjunction with the CNN-DM corpus (Hermann et al., 2015) as our designated testing dataset.
Dataset Splits No The paper mentions "1,999 examples in the test set" for WMT 14 and "11,490 examples" for CNN-DM (implied test set here) but doesn't explicitly state the train/validation splits or how these datasets were partitioned for those purposes. No explicit validation split information or full dataset splits are provided.
Hardware Specification Yes All experiments are conducted on three Nvidia A6000 GPUs with 48GB of memory.
Software Dependencies No The experiments are implemented using the Huggingface library (Wolf et al., 2019), a widely adopted platform for model development and sharing within the NLP community. No specific version numbers are provided for Huggingface or other key software dependencies.
Experiment Setup Yes Specifically for Di Pmark, we select values for α from the set {0.3, 0.35, 0.4, 0.45, 0.5}, while for the Soft watermark (Kirchenbauer et al., 2023), we choose green list bias values δ from the set {0.0, 1.0, 1.5, 2.0} alongside a fixed green list separator γ = 0.5, indicating that 50% of tokens are green while the remainder are red. Texture key generation relies on the most recent five tokens as texture key.