Duwak

Welcome to the official repository for the paper "Duwak: Dual Watermarks in Large Language Models."

Installation

To install the necessary dependencies, run the following command:

./install.sh

Usage

To use Duwak for embedding dual watermarks in large language models, follow these steps:

1. Prepare Watermark Specifications:

   • Create a file with the specifications for the watermarks you wish to embed. Each line should contain a JSON object that describes the parameters of a watermark.
   • Example watermark specifications can be found in ./run/watermark_specs.
   • For a complete definition of the WatermarkSpec object, refer to ./src/watermark-benchmark/utils/classes.py.

2. Configure Settings:

   • Set up your configuration file based on the provided example in ./run/config.yml.

3. Set Python Path:

   • Include the watermark_benchmark in your Python path with the following command:

     export PYTHONPATH="path_to_src:$PYTHONPATH"

4. Run Benchmark:

   • Execute the benchmark script with the following command:

     cd run
     python3 -m ../src/watermark-benchmark/pipeline/run_all.py config.yml

Our Watermarks

• Duwak watermarking scheme is implemented in ./src/watermark-benchmark/watermark/schemes/contrastive_distri.py.
• This codebase is an extension of the MarkMyWords project. The original codebase is available at MarkMyWords GitHub Repository.

Docker Support

Build Image (Optional):

To build the Docker image from the source code, use the following command:

docker build -t duwak .

Run Container:

To run the Docker container with your source code, replace /path_to_your_source with the actual path to your source code. Use the following command:

cd run
docker run -it --rm \\
    --gpus all \\
    -v /path_to_your_src:/app/src \\
    -v $(pwd):/project \\
    -w /project \\
    -e PYTHONPATH="/app/src:$PYTHONPATH" \\
    -e HF_TOKEN=your_HF_TOKEN \\
    duwak \\
    bash -c "python3 ./src/watermark_benchmark/pipeline/run_all.py ./run/config.yml"

This setup allows you to easily use the duwak image for Docker container operations.