Synthetic Data SDK
The Synthetic Data SDK by MOSTLY AI is an open-source Python toolkit designed for data scientists, engineers, and privacy officers who need high-fidelity, privacy-safe synthetic data.
Publisher review
The Synthetic Data SDK by MOSTLY AI is an open-source Python toolkit designed for data scientists, engineers, and privacy officers who need high-fidelity, privacy-safe synthetic data. It targets users working with mixed-type tabular data—including categorical, numerical, geospatial, and text fields—who want to generate synthetic copies for testing, analytics, or machine learning without exposing sensitive records. The SDK operates in two modes: LOCAL, which trains and generates data on the user's own compute (CPU or GPU), and CLIENT, which connects to the MOSTLY AI Data Intelligence Platform for cloud-based training and generation. It is freely available under a permissive license, making it accessible for individual developers and small teams, while enterprise-grade features require paid plans.
The SDK provides three core resource types: Generators (trained on existing tabular or language data), Synthetic Datasets (produced by sampling from a generator), and Connectors (for reading/writing data from databases or cloud storage). It supports multiple model architectures: TabularARGN for state-of-the-art tabular performance, fine-tuned HuggingFace language models for text-rich data, and an efficient LSTM for text synthesis from scratch. Advanced training options include differential privacy (to bound re-identification risk), real-time progress monitoring, and automated quality assurance with built-in fidelity and privacy metrics. Sampling is flexible: users can up-sample to any volume (e.g., 1 million rows), conditionally generate by any column, rebalance underrepresented segments, impute missing values with context awareness, enforce statistical fairness controls, and adjust rule adherence via a temperature parameter. The SDK also generates in-depth HTML reports for visual analysis of synthetic data quality.
MOSTLY AI competes directly with Gretel.ai and K2view in the synthetic data generation market. Gretel.ai offers a similar open-source SDK with cloud-based APIs, while K2view provides a broader data integration platform that includes synthetic data capabilities. MOSTLY AI was recognized as a Visionary in Gartner’s Magic Quadrant for Data Integration Tools, indicating its strategic positioning. The SDK’s evaluation framework, co-developed with academic research, provides empirical methods to quantify both accuracy (utility) and privacy—a differentiator that helps users compare generators on objective metrics. However, the open-source SDK lacks some enterprise features (e.g., centralized governance, team collaboration) that are only available in the paid platform version.
Key trade-offs: The SDK’s LOCAL mode is compute-intensive—training a TabularARGN model on a dataset with 100,000 rows and 50 columns can take hours on a CPU, and GPU acceleration is recommended but not free. Differential privacy reduces model fidelity; a privacy budget (epsilon) of 1.0 may degrade utility by 15-30% depending on the dataset. The SDK does not natively support streaming or real-time data generation—it is batch-oriented. Finally, while the open-source license is permissive, the most advanced features (e.g., multi-table synthesis, automated data catalog integration) require the paid MOSTLY AI Platform, limiting the free tier’s capability for complex enterprise use cases.
How it works
-
Broad data type support
Handles categorical, numerical, geospatial, text, and mixed-type data in single-table, multi-table, and time-series formats.
-
Multiple model architectures
Includes TabularARGN for SOTA tabular performance, fine-tuned HuggingFace models for text, and efficient LSTM for text synthesis.
-
Advanced training options
Supports GPU/CPU training, differential privacy with configurable epsilon, and real-time progress monitoring via logs.
-
Automated quality assurance
Built-in fidelity and privacy metrics generate in-depth HTML reports for visual analysis of synthetic data quality.
-
Flexible sampling methods
Allows up-sampling to any volume, conditional generation by any column, rebalancing segments, context-aware imputation, and fairness controls.
-
External data connectors
Connects to databases and cloud storage for reading/writing data, enabling integration with existing data pipelines.
-
Open-source permissive license
Freely available under a permissive open-source license, allowing modification and redistribution without vendor lock-in.
Strengths and trade-offs
Strengths
- Provides a research-backed evaluation framework with empirical metrics for fidelity and privacy, enabling objective benchmarking against other generators.
- Supports diverse data types (categorical, numerical, geospatial, text) and complex structures (multi-table, time-series) in a single SDK.
- Offers advanced training options including differential privacy with configurable epsilon and real-time progress monitoring via logs.
- Integrates with external data sources (databases, cloud storage) and includes connectors for reading/writing data programmatically.
Trade-offs
- Training TabularARGN models on large datasets (100k+ rows) can take hours on CPU; GPU acceleration is recommended but adds hardware cost.
- Differential privacy reduces model fidelity—a privacy budget of epsilon=1.0 may degrade utility by 15-30% depending on dataset characteristics.
- The SDK is batch-oriented and does not natively support streaming or real-time synthetic data generation for live applications.
- Advanced enterprise features (multi-table synthesis, automated catalog integration) require the paid MOSTLY AI Platform, limiting the free tier's capability.
Pricing context
Free open-source version available under a permissive license; paid enterprise plans for advanced features (pricing not publicly disclosed).
Getting started with Synthetic Data SDK
-
Install the SDK
Install the Synthetic Data SDK via pip: `pip install mostly-ai`. Ensure Python 3.8+ is available. For GPU acceleration, install CUDA-compatible PyTorch separately.
-
Connect to data source
Use a Connector to read your tabular data. For a CSV file, call `mostly.create_connector(type='file', file_path='data.csv')`. For databases, specify connection parameters.
-
Configure a generator
Create a Generator object with `mostly.create_generator(name='my_generator')`. Set training parameters like model architecture (e.g., `model='tabular_argn'`) and differential privacy epsilon if needed.
-
Train and generate data
Train the generator on your data: `generator.fit(data)`. Then sample synthetic data: `synthetic_data = generator.generate(num_rows=10000)`. Adjust sampling parameters for conditional generation or rebalancing.
-
Evaluate and export results
Run quality assurance with `report = mostly.report(synthetic_data, original_data)`. Export the HTML report for visual analysis. Save synthetic data to CSV or push to a database via a Connector.
Frequently Asked Questions
What is the Synthetic Data SDK by MOSTLY AI?
The Synthetic Data SDK by MOSTLY AI is an open-source Python toolkit for generating high-fidelity, privacy-safe synthetic data from mixed-type tabular data. It targets data scientists, engineers, and privacy officers needing synthetic copies for testing, analytics, or machine learning without exposing sensitive records.
How does the Synthetic Data SDK handle different data types?
The SDK supports categorical, numerical, geospatial, and text fields in single-table, multi-table, and time-series formats. It uses model architectures like TabularARGN for tabular data, fine-tuned HuggingFace models for text, and an efficient LSTM for text synthesis from scratch.
What are the training modes available in the Synthetic Data SDK?
The SDK operates in two modes: LOCAL, which trains and generates data on the user's own compute using CPU or GPU, and CLIENT, which connects to the MOSTLY AI Data Intelligence Platform for cloud-based training and generation. LOCAL mode is compute-intensive for large datasets.
Does the Synthetic Data SDK support differential privacy?
Yes, the SDK offers differential privacy with a configurable epsilon budget to bound re-identification risk. However, a privacy budget of epsilon=1.0 may degrade model fidelity by 15-30% depending on the dataset, reducing utility for some use cases.
What sampling options does the Synthetic Data SDK provide?
The SDK allows flexible sampling including up-sampling to any volume, conditional generation by any column, rebalancing underrepresented segments, context-aware imputation of missing values, statistical fairness controls, and adjustment of rule adherence via a temperature parameter.
What are the limitations of the open-source Synthetic Data SDK?
The open-source SDK lacks enterprise features like centralized governance and team collaboration, which require the paid MOSTLY AI Platform. It is batch-oriented without streaming support, and training large models on CPU can take hours, with GPU acceleration recommended but not free.
Alternatives
How Synthetic Data SDK compares
Direct head-to-head against 2 competitors. Picked by 7wData.
Synthetic Data SDK
- Pricing
- Free open-source version available under a permissive license; paid enterprise plans for advanced features (pricing not publicly disclosed).
- Target
- The Synthetic Data SDK by MOSTLY AI is an open-source Python toolkit designed for data scientists, engineers, and privacy officers who need high-fidelity, privacy-safe synthetic
- Strength
- Provides a research-backed evaluation framework with empirical metrics for fidelity and privacy, enabling objective benchmarking against other generators.
- Watch for
- Training TabularARGN models on large datasets (100k+ rows) can take hours on CPU; GPU acceleration is recommended but adds hardware cost.
Tonic Fabricate
- Pricing
- Custom/Contact sales
- Target
- Developers needing AI-generated relational datasets
- Deployment
- Cloud/SaaS
- Strength
- Agentic approach with chat-based data generation
- Watch for
- Pricing opacity for enterprise plans
MOSTLY AI
- Pricing
- Open-source core, enterprise pricing undisclosed
- Target
- Teams requiring privacy-safe synthetic data
- Deployment
- Local or cloud
- Strength
- Local-first SDK with platform integration
- Watch for
- Limited unstructured data support
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.