synthid-text library
Reference implementation of the text watermark for Hugging Face Transformers. Useful for anyone who wants to watermark their own model's output, or to study how the signal behaves after editing.
What it is
The library wraps the watermarking logic as a logits processor for Hugging Face Transformers. You supply a key, generate text through any compatible model, and later score passages with the matching detector. It is the simplest way to see, with your own eyes, how the signal rises with passage length and falls after paraphrasing. Researchers use it as the reference implementation when they test attacks.
When to use this entry
Use it to experiment, to watermark a model you deploy, or to reproduce a paper's robustness numbers.
Questions
Which models does it work with?
Any causal language model served through Transformers that exposes logits during sampling. It was released alongside Gemma examples.
Is it the same code Google runs in Gemini?
It is the published method; Google's production system and keys are separate.