Gretel Synthetics: Synthetic Data Generation for Structured and Unstructured Text
Gretel Synthetics was a synthetic data generation library developed by Gretel.ai, now archived after Gretel’s acquisition by NVIDIA in 2025. The library focused on generating synthetic data for structured and unstructured text, with optional differential privacy to enhance data anonymization. Its core models included ACTGAN, DGAN for time series, and LSTM-based approaches. While Gretel Synthetics itself is no longer actively maintained, its foundational work influenced later tools, such as NVIDIA’s NeMo Data Designer, which continues Gretel’s data-centric AI initiatives.
The library was designed to address privacy concerns in machine learning by enabling organizations to create synthetic datasets that mimic real-world data distributions without exposing sensitive information. This was particularly valuable in industries where data privacy regulations, such as GDPR, impose strict constraints on the use of real user data.
Gretel Synthetics was released under a custom or multiple licensing model, depending on the use case, and was primarily implemented in Python. Its development spanned from 2020 until its archival in 2025, during which it accumulated 683 stars on GitHub, reflecting its adoption within the AI and data science communities.
The project’s repository served as a central hub for documentation, examples, and model implementations, making it accessible for researchers and practitioners seeking to generate synthetic data for training and testing machine learning models.
How It Works
Gretel Synthetics provided a modular framework for generating synthetic data, supporting three primary model architectures: ACTGAN, DGAN for time series, and LSTM-based models. Each model was optimized for specific data types and use cases, ensuring flexibility in synthetic data generation.
The library’s architecture relied on external dependencies to ensure compatibility and performance. For instance, PyTorch was required for the Timeseries DGAN and ACTGAN models, with specific version recommendations to avoid conflicts. SDV (Synthetic Data Vault) was also a key dependency for ACTGAN, as it provided additional functionality such as column autodetection and memory optimization. Users were advised to install these dependencies separately to align with the library’s requirements.
Installation of Gretel Synthetics involved several steps, including cloning the repository and installing core dependencies. The library supported both CPU and GPU environments, with a dedicated setup script for GPU acceleration on Ubuntu 18.04. Users could also install the package directly via pip, though additional configurations were necessary for full functionality.
These examples covered structured data, time series, and unstructured text, offering practical guidance for implementation.
Timeseries DGAN
The Timeseries DGAN module implemented a PyTorch-based DoppelGANger model tailored for time series data. This model was designed to capture temporal dependencies in sequential data, making it suitable for applications such as smart home sensor readings, financial forecasting, or IoT device analytics. The implementation required PyTorch version 1.13.1, and users were directed to a dedicated notebook for basic usage examples.
This made it particularly useful for scenarios where real-time data was sensitive or where synthetic data could replace live feeds for testing and validation purposes.
ACTGAN (Anyway CTGAN)
ACTGAN was an extension of the Conditional Tabular GAN (CTGAN) model, incorporating enhancements such as improved memory management, automated column transformation, and better handling of mixed data types. Unlike the standard CTGAN, ACTGAN included additional features to streamline the synthetic data generation process, such as reduced manual configuration requirements.
The ACTGAN interface was backward-compatible with CTGAN, meaning users familiar with the original model could transition smoothly to ACTGAN with minimal adjustments.
ACTGAN relied on SDV (Synthetic Data Vault) for core functionality, which introduced potential dependency conflicts if users had previously installed specific versions of PyTorch for other models like Timeseries DGAN. To mitigate this, the library recommended installing SDV separately and ensuring version compatibility.
Use Cases
Gretel Synthetics was primarily used in scenarios where synthetic data could replace real data to mitigate privacy risks.
- Privacy-Preserving Machine Learning: Generating synthetic datasets for training models without exposing raw user data, ensuring compliance with regulations like GDPR or HIPAA.
- Time Series Analysis: Generating synthetic time series data for applications in finance, healthcare, or industrial IoT, where temporal patterns are critical.
The library’s support for differential privacy further enhanced its utility in high-security environments, where even synthetic data required additional safeguards against re-identification.
Key Features and Capabilities
- Model Diversity: Support for multiple architectures (ACTGAN, DGAN, and LSTM) allowed users to select the most appropriate model for their data type and use case.
- Differential Privacy: Optional integration of differential privacy techniques ensured that synthetic data could not be traced back to original records, adding an extra layer of security.
- GPU Acceleration: The library provided optimized setup scripts for GPU-enabled environments, significantly reducing generation times for large datasets.
- Versioned Dependencies: Clear recommendations for dependency versions (e.g., PyTorch 2.0 for ACTGAN, SDV 0.17.x) ensured reproducibility and compatibility across different environments.
Technical Evolution and Archival
Gretel Synthetics underwent several iterations between 2020 and 2025, with each release introducing fixes, dependency updates, and architectural refinements.
These updates reflected Gretel’s commitment to maintaining a stable and efficient toolkit, even as the company transitioned its focus toward NVIDIA’s broader AI ecosystem. The archival of Gretel Synthetics in 2025 did not render its contributions obsolete; instead, its core principles and models influenced subsequent tools like NVIDIA NeMo Data Designer, which continues to build on Gretel’s legacy in synthetic data generation.