Text Seal

Meta Text Seal is a comprehensive toolkit for LLM generation-time watermarking, post-hoc text watermarking through LLM rephrasing, and contamination detection through watermark radioactivity. It is part of the Meta Seal family of watermarking technologies.

Features

• 🔏 Post-hoc Watermarking: Rephrase text with an LLM while inserting a watermark using generation-time scheme (Green-list/Red-list, Gumbel-max, DipMark, SynthID, MorphMark, WaterMax).
• 🧪 Contamination Detection: Detect watermarked dataset membership inference through radioactivity.
• 🚀 Training Infrastructure: Distributed pretraining and SFT with contamination injection support for research purposes.

Papers

This codebase implements methods from:

• How Good is Post-Hoc Watermarking With Language Model Rephrasing?: Post-hoc watermarking through rephrasing with a watermarked LLM.

• Detecting Benchmark Contamination Through Watermarking: Detecting training data contamination with watermarked benchmarks.

Quick Start

Installation

# Clone the repository
git clone https://github.com/facebookresearch/textseal.git
cd textseal

# Create environment and install dependencies
conda create -n text_seal python=3.11.13
conda activate text_seal
pip install -r requirements.txt

💡 For contamination detection experiments (training with contamination injection), you need additional setup. First follow the Meta Lingua installation instructions, then install the requirements above. See Environment Setup for details.

Post-hoc Watermarking

Watermark existing text by rephrasing with an LLM (here using Gumbel-max watermarking and Llama-3.2-3B-Instruct):

python -m apps.posthoc.main \
  --input_path assets/sample_document.txt \
  --dump_dir output/ \
  --watermark.watermark_type gumbelmax \
  --model.model_name meta-llama/Llama-3.2-3B-Instruct \
  --processing.temperature 1.0 \
  --processing.top_p 0.95

Results are saved in output/ directory as a JSONL file containing original, watermarked text and statistics.

See docs/README_posthoc.md for detailed documentation.

Contamination Detection

Inject watermarked benchmarks during training and detect memorization through watermark radioactivity.

Download DCLM training data and benchmark datasets (ARC-Easy, ARC-Challenge, MMLU). See Data Preparation in the contamination docs.

The contamination detection workflow consists of three steps, each with its own experiment configuration file:

# Step 1: Watermark benchmarks with different secret keys
python -m apps.posthoc.main --config configs/watermark_benchmarks.yaml

# Step 2: Train model with contaminated watermarked data
python -m apps.common.stool script=apps.wmtraining.train \
  config=configs/train_with_contamination.yaml \
  nodes=4 ngpu=8 partition=learn qos=high time=4320

# Step 3: Detect contamination via watermark evaluation
python -m apps.wmtraining.eval_wm --config configs/eval_contamination.yaml

Configuration files:

• configs/watermark_benchmarks.yaml - Watermark benchmark datasets
• configs/train_with_contamination.yaml - Train with contamination injection
• configs/eval_contamination.yaml - Evaluate contamination detection

See docs/README_contamination.md for detailed documentation.

Documentation

• Post-hoc Watermarking - Rephrase text while adding a watermark
• Contamination Detection - Detect benchmark memorization via watermarks

Repository Structure

textseal/
├── apps/
│   ├── posthoc/          # Post-hoc watermarking
│   ├── wmtraining/       # Training and evaluation
│   ├── analysis/         # Analysis tools
│   └── common/           # Shared utilities (LLM, watermark, config)
├── docs/                 # Detailed documentation
├── configs/              # Example configurations for watermarking and training
├── assets/               # Sample texts
├── setup/                # Setup scripts and data processing

Use Cases

1. Content Authentication

Watermark text to enable verification and provenance tracking.

2. Dataset Contamination Detection

Detect if evaluation benchmarks were included in training data by injecting watermarked versions and checking for "radioactivity."

3. Research on Watermarking

Experiment with different watermarking algorithms and detection methods on your own models and datasets.

License

Meta Text Seal is released under the MIT License.

It relies on code and models from other repositories. The contamination detection app builds on Meta Lingua for training, which has a BSD 3-Clause License. The models used for post-hoc watermarking are loaded from Hugging Face and are subject to their respective licenses.

Citation

If you use Text Seal in your research, please cite:

@article{sander2025detecting,
  title={Detecting benchmark contamination through watermarking},
  author={Sander, Tom and Fernandez, Pierre and Mahloujifar, Saeed and Durmus, Alain and Guo, Chuan},
  journal={arXiv preprint arXiv:2502.17259},
  year={2025}
}

@article{fernandez2025how,
  title={How Good is Post-Hoc Watermarking With Language Model Rephrasing?},
  author = {Fernandez, Pierre and Sander, Tom and Elsahar, Hady and Chang, Hongyan and Sou\v{c}ek, Tom\'{a}\v{s} and Lacatusu, Valeriu and Tran, Tuan and Rebuffi, Sylvestre-Alvise and Mourachko, Alexandre},
  year={2025}
}