Robust Distortion-free Watermarks for Language Models

Implementation of the methods described in Robust Distortion-free Watermarks for Language Models.

by Rohith Kuditipudi, John Thickstun, Tatsunori Hashimoto, and Percy Liang.

---

This repository provides code that implements the watermarks described in Robust Distortion-free Watermarks for Language Models. See also the blog post, which includes an in-browser demo of the watermark detector.

Setup

We provide a list of package requirements in requirements.txt, from which you can create a conda environment by running

conda create --name <env> --file requirements.txt

We recommend installing version 4.30.1 or earlier of the transformers package, as otherwise the roundtrip translation experiments may break. Also, be sure to set the environment variables $HF_HOME and $TRANSFORMERS_CACHE to a directory with sufficient disk space.

Basic Usage

We provide standalone Python code for generating and detecting text with a watermark, using our recommended instantiation of the watermarking strategies discussed in the paper in demo/generate.py and demo/detect.py. We also provide the Javascript implementation of the detector demo/detect.js used for the in-browser demo.

To generate m tokens of text from a model (e.g., facebook/opt-1.3b) with watermark key 42, run:

python demo/generate.py --model facebook/opt-1.3b --m 80 --key 42 > doc.txt

Checking for the watermark requires a watermark key (in this case, 42) and the model tokenizer, but crucially it does not require access to the model itself. To test for a watermark in a given text document doc.txt, run

python demo/detect.py doc.txt --tokenizer facebook/opt-1.3b --key 42

Alternatively, you can use the javascript detector implemented in demo/detect.js which runs much faster (this is also the detector used for the web demo).

Reproducing experiments from the paper

The experiments directory contains code for reproducing the experimental results we report in the paper (in particular, Experiments 1-7 as described in Section 3). We use experiments/c4-experiment.py to run the news-like C4 dataset experiments (Experiments 1-6) and experiments/instruct-experiment.py to run the Alpaca instruction evaluation dataset experiments (Experiment 7). In particular,

python experiments/{c4,instruct}-experiment.py --save results.p

will save experiment settings and results as a Python dict. See experiments/analyze.py for a minimal example of how to parse this dict.

We include shell scripts for reproducing all experiments in experiments/scripts. You will need to specify certain experiment settings in order to run most of the scripts. For example, to reproduce Figure 2a in the paper using the OPT-1.3B model, run

./experiments/scripts/experiment-1.sh <save directory path> facebook/opt-1.3b 

To reproduce Figure 4b in the paper using the LLaMA-7B model with m = 35, run

./experiments/scripts/experiment-3.sh arxiv-results/experiment-3/llama huggyllama/llama-7b 35

And as a final example, to reproduce Figure 9a in the paper using the LLaMA-7B model and with French as the roundtrip language, run

./experiments/scripts/experiment-6.sh arxiv-results/experiment-6/opt facebook/opt-1.3b french

---

@article{kuditipudi2023robust,
  title={Robust Distortion-free Watermarks for Language Models},
  author={Kuditipudi, Rohith and Thickstun, John and Hashimoto, Tatsunori and Liang, Percy},
  journal={arXiv preprint arXiv:2307.15593},
  year={2023}
}