google-deepmind/synthid-text ? reverse-engineered prompt

Build me a Python library and notebook project that shows how to watermark text generated by a language model and then detect that watermark afterward.

I want it to work as a reference implementation for research, not production. It should include the core watermarking logic, a simple way to plug into Hugging Face text generation models like Gemma and GPT 2, and a few detection options, including a lightweight score based detector and a stronger trained detector. Please make the notebook easy to run end to end with sample inputs, and include clear examples for generating watermarked text and scoring whether text was likely watermarked.

Also add a small test suite so I can verify everything still works, plus setup instructions for installing from source in a virtual environment and running the notebook locally. If you need to check current library docs online while wiring up the model integration, go ahead and do that.