1. Upload the code.

• Install
• Dataset
• Generation and Detection
• Attacks

conda create -n pmark python=3.10
conda activate pmark
pip install -r requirements.txt

We have upload a demo of C4 and BOOKSUM dataset in folder ./data.

To generate watermarked text with PMark, run

bash scripts/gen.sh 

Also, to get help about arguments in pmark.py or detect.py, run

python pmark.py -h
python detect.py -h

We use standard attack methods implemented in MarkLLM(https://github.com/THU-BPM/MarkLLM), defined in attack_utils.py. Therefore to conduct attack experiments, just put attack_utils.py and attack_pmark.py in the root dir of MarkLLM repo and run

bash scripts/attack.sh