brianpulfer / lmwatermark

Implementation of 'A Watermark for Large Language Models' paper by Kirchenbauer & Geiping et. al.

Hey! 👋

Thanks for coming by. My name is Brian Pulfer and I am currently a Ph.D student in Machine Learning at the University of Geneva, Switzerland 🇨🇭.

I am fascinated by the elegance and power of Machine Learning and I am currently learning as much as possible to maximize my positive impact on humanity.

Useful Links 🔗

🌐 Website 🐦 Twitter, 👔 LinekdIn

lmwatermark's Issues

Salted Hash

Hi, great implementation :)

A minor comment, I think you are using the default python hash. This hash function is salted by default, see https://docs.python.org/3/reference/datamodel.html#object.__hash__. A watermarked text fragment generated in this way, and saved to file, would later not be detectable in a new process, as the salt is unknown and python process-specific.