Unigram-Watermark

This repository contains the code for the paper Provable Robust Watermarking for AI-Generated Text.

🔍 HuggingFace Demo

If you find this repository useful, please cite our paper:

@article{zhao2023provable,
  title={Provable Robust Watermarking for AI-Generated Text},
  author={Zhao, Xuandong and Ananth, Prabhanjan and Li, Lei and Wang, Yu-Xiang},
  journal={arXiv preprint arXiv:2306.17439},
  year={2023}
}

Example

First, download the data by running the following command:

cd data
bash download_data.sh

You can generate watermarked text by running the following command:

python run_generate.py

You can detect the watermark by running the following command:

python run_detect.py --input_file {your_input_file}

Acknowledgement

We thank the authors of the following research works and open-source projects:

A Watermark for Large Language Models

Paraphrasing evades detectors of AI-generated text, but retrieval is an effective defense