deepmark 1.0.0

deepmark: Model-agnostic, robust, and compliance-ready text watermarking library for LLMs and content platforms.

deepmark Python SDK

Overview

A model-agnostic, robust, and compliance-ready text watermarking library for LLMs and content platforms. Supports both generation-time and post-processing watermark embedding, with detection tools for developers.

Features

• Generation-time and post-generation watermarking
• Model-agnostic (OpenAI, HuggingFace, custom LLMs)
• Robust to paraphrasing, translation, and editing
• Python SDK for easy integration
• Batch detection, confidence scoring
• Multilingual support (English, Spanish, French, German, Italian, Portuguese)

Installation

pip install -r requirements.txt

Usage

Marking Text

from deepmark import mark_text

# Mark text with a watermark
marked = mark_text("Hello world", key="your-secret-key", source_id="GPT-4")
print(marked)

Detecting Watermarks

from deepmark import detect_watermark

# Detect watermark in text
result = detect_watermark(marked)
print(result)
# Output: {'confidence_score': ..., 'source_id': ..., 'timestamp': ..., 'key_id': ..., 'tampering_likelihood': ...}

Extensibility

• Add more languages by downloading spaCy models and updating synonym logic.
• Add more metadata by extending the zero-width encoding.
• Integrate with LLMs or pipelines by calling mark_text and detect_watermark.

License

MIT