SynthID Text
This repository provides a reference implementation of the SynthID Text watermarking and detection capabilities for the research paper published in Nature. It is not intended for production use. The core library is distributed on PyPI for easy installation in the Python Notebook example, which demonstrates how to apply these tools with the Gemma and GPT-2 models.
Installation and usage
The Colab Notebook is self-contained reference implementation that:
- Extends the
GemmaForCausalLMandGPT2LMHeadModelclasses from Hugging Face Transformers with a mix-in to enable watermarking text content generated by models running in PyTorch; and - Detects the watermark. This can be done either with the simple Weighted Mean detector which requires no training, or with the more powerful Bayesian detector that requires training. If using the Weighted Mean detector approach across texts of varying token lengths, we recommend empirically/theoretically computing the thresholds at the desired false positives rate at specific token lengths, or using a weighted frequentist approach as described in Appendix A.3.1.
The notebook is designed to be run end-to-end with either a Gemma or GPT-2 model, and runs best on the following runtime hardware, some of which may require a Colab Subscription.
- Gemma v1.0 2B IT: Use a GPU with 16GB of memory, such as a T4.
- Gemma v1.0 7B IT: Use a GPU with 32GB of memory, such as an A100.
- GPT-2: Any runtime will work, though a High-RAM CPU or any GPU will be faster.
NOTE: This implementation is for reference and research reproducibility purposes only. Due to minor variations in Gemma and Mistral models across implementations, we expect minor fluctuations in the detectability and perplexity results obtained from this repository versus those reported in the paper. The subclasses introduced herein are not designed to be used in production systems. Check out the official SynthID Text implementation in Hugging Face Transformers for a production-ready implementation.
NOTE: The synthid_text.hashing_function.accumulate_hash() function, used while
computing G values in this reference implementation, does not provide any
guarantees of cryptographic security.
Local notebook use
The notebook can also be used locally if installed from source. Using a virtual environment is highly recommended for any local use.
# Create and activate the virtual environment python3 -m venv ~/.venvs/synthid source ~/.venvs/synthid/bin/activate # Download and install SynthID Text and Jupyter git clone https://github.com/google-deepmind/synthid-text.git cd synthid-text pip install '.[notebook-local]' # Start the Jupyter server python -m notebook
Once your kernel is running navigate to .pynb file to execute.
Running the tests
The source installation also includes a small test suite to verify that the library is working as expected.
# Create and activate the virtual environment python3 -m venv ~/.venvs/synthid source ~/.venvs/synthid/bin/activate # Download and install SynthID Text with test dependencies from source git clone https://github.com/google-deepmind/synthid-text.git cd synthid-text pip install '.[test]' # Run the test suite pytest .
How it works
Defining a watermark configuration
SynthID Text produces unique watermarks given a configuration, with the most
important piece of these configurations being the keys: a sequence of unique
integers where len(keys) corresponds to the number of layers in the
watermarking or detection models.
The structure of a configuration is described in the following TypedDict
subclass, though in practice, the mixin class in this library
uses a static configuration.
from collections.abc import Sequence from typing import TypedDict import torch class WatermarkingConfig(TypedDict): ngram_len: int keys: Sequence[int] sampling_table_size: int sampling_table_seed: int context_history_size: int device: torch.device
Applying a watermark
Watermarks are applied by a mix-in class that wraps the
GemmaForCausalLM and
GPT2LMHeadModel classes from Transformers, which results
in two subclasses with the same API that you are used to from Transformers.
Remember that the mix-in provided by this library uses a static watermarking
configuration, making it unsuitable for production use.
from synthid_text import synthid_mixin import transformers import torch DEVICE = ( torch.device('cuda:0') if torch.cuda.is_available() else torch.device('cpu') ) INPUTS = [ "I enjoy walking with my cute dog", "I am from New York", "The test was not so very hard after all", "I don't think they can score twice in so short a time", ] MODEL_NAME = 'google/gemma-2b-it' TEMPERATURE = 0.5 TOP_K = 40 TOP_P = 0.99 # Initialize a standard tokenizer from Transformers. tokenizer = transformers.AutoTokenizer.from_pretrained(MODEL_NAME) # Initialize a SynthID Text-enabled model. model = synthid_mixin.SynthIDGemmaForCausalLM.from_pretrained( MODEL_NAME, device_map='auto', torch_dtype=torch.bfloat16, ) # Prepare your inputs in the usual way. inputs = tokenizer( INPUTS, return_tensors='pt', padding=True, ).to(DEVICE) # Generate watermarked text. outputs = model.generate( **inputs, do_sample=True, max_length=1024, temperature=TEMPERATURE, top_k=TOP_K, top_p=TOP_P, )
Detecting a watermark
Watermark detection can be done using a variety of scoring functions (see paper). This repository contains code for the Mean, Weighted Mean, and Bayesian scoring functions described in the paper. The colab contains examples for how to use these scoring functions.
The Bayesian detector must be trained on watermarked and unwatermarked data before it can be used. The Bayesian detector must be trained for each unique watermarking key, and the training data used for this detector model should be independent from, but representative of the expected character and quality of the text content the system will generate in production.
import jax.numpy as jnp from synthid_text import train_detector_bayesian def load_data(): # Get your training and test data into the system. pass def process_training_data(split): # Get the G values, masks, and labels for the provided split. pass train_split, test_split = load_data() train_g_values, train_masks, train_labels = process_training_data(train_split) test_g_values, test_masks, test_labels = process_training_data(test_split) detector, loss = train_detector_bayesian.optimize_model( jnp.squeeze(train_g_values), jnp.squeeze(train_masks), jnp.squeeze(train_labels), jnp.squeeze(test_g_values), jnp.squeeze(test_masks), jnp.squeeze(test_labels), )
Once the Bayesian detector is trained, use the detector.score() function to
generate a per-example score indicating if the text was generated with the given
watermarking configuration. Score values will be between 0 and 1, with scores
closer to 1 indicating higher likelihood that the text was generated with the
given watermark. You can adjust the acceptance threshold to your needs.
from synthid_text import logits_processing CONFIG = synthid_mixin.DEFAULT_WATERMARKING_CONFIG logits_processor = logits_processing.SynthIDLogitsProcessor( **CONFIG, top_k=TOP_K, temperature=TEMPERATURE ) # Get only the generated text from the models predictions. outputs = outputs[:, inputs_len:] # Compute the end-of-sequence mask, skipping first ngram_len - 1 tokens # <bool>[batch_size, output_len] eos_token_mask = logits_processor.compute_eos_token_mask( input_ids=outputs, eos_token_id=tokenizer.eos_token_id, )[:, CONFIG['ngram_len'] - 1 :] # Compute the context repetition mask # <bool>[batch_size, output_len - (ngram_len - 1)] context_repetition_mask = logits_processor.compute_context_repetition_mask( input_ids=outputs ) # Compute the mask that isolates the generated text. combined_mask = context_repetition_mask * eos_token_mask # Compute the G values for the generated text. g_values = logits_processor.compute_g_values(input_ids=outputs) # Score the G values, given the combined mask, and output a per-example score # indicating whether the detector.score(g_values.cpu().numpy(), combined_mask.cpu().numpy())
Human Data
We release the human evaluation data, where we compare watermarked text against unwatermarked text generated from the Gemma 7B model.
The data is located in data/human_eval.jsonl.
To get the prompts used for generating the responses, please use the following code.
import json import tensorflow_datasets as tfds ds = tfds.load('huggingface:eli5/LFQA_reddit', split='test_eli5') id_to_prompt = {} for x in ds.as_numpy_iterator(): id_to_prompt[x['q_id'].decode()] = x['title'].decode() full_data = [] with open('./data/human_eval.jsonl') as f: for json_str in f: x = json.loads(json_str) x['question'] = id_to_prompt[x['q_id']] full_data.append(x)
Citing this work
@article{Dathathri2024, author={Dathathri, Sumanth and See, Abigail and Ghaisas, Sumedh and Huang, Po-Sen and McAdam, Rob and Welbl, Johannes and Bachani, Vandana and Kaskasoli, Alex and Stanforth, Robert and Matejovicova, Tatiana and Hayes, Jamie and Vyas, Nidhi and Merey, Majd Al and Brown-Cohen, Jonah and Bunel, Rudy and Balle, Borja and Cemgil, Taylan and Ahmed, Zahra and Stacpoole, Kitty and Shumailov, Ilia and Baetu, Ciprian and Gowal, Sven and Hassabis, Demis and Kohli, Pushmeet}, title={Scalable watermarking for identifying large language model outputs}, journal={Nature}, year={2024}, month={Oct}, day={01}, volume={634}, number={8035}, pages={818-823}, issn={1476-4687}, doi={10.1038/s41586-024-08025-4}, url={https://doi.org/10.1038/s41586-024-08025-4} }
License and disclaimer
Copyright 2024 DeepMind Technologies Limited
All software is licensed under the Apache License, Version 2.0 (Apache 2.0); you may not use this file except in compliance with the Apache 2.0 license. You may obtain a copy of the Apache 2.0 license at: https://www.apache.org/licenses/LICENSE-2.0
All other materials are licensed under the Creative Commons Attribution 4.0 International License (CC-BY). You may obtain a copy of the CC-BY license at: https://creativecommons.org/licenses/by/4.0/legalcode
Unless required by applicable law or agreed to in writing, all software and materials distributed here under the Apache 2.0 or CC-BY licenses are distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the licenses for the specific language governing permissions and limitations under those licenses.
This is not an official Google product.