This Python SDK provides a toolkit for generating high-fidelity, privacy-safe synthetic tabular and language data. It targets data scientists and engineers needing to create realistic datasets for testing, development, or privacy-preserving analytics, offering both local and cloud-based generation capabilities.

How It Works

The SDK utilizes a modular approach, allowing users to train generators on their data, create synthetic datasets from these generators, and connect to various data sources. It features state-of-the-art models like TabularARGN for tabular data, offering significant efficiency gains, and supports fine-tuning Hugging Face language models or using custom LSTMs for text synthesis.

Quick Start & Requirements

• Install via pip: uv pip install -U 'mostlyai[local-gpu]' (recommended for GPU) or uv pip install -U mostlyai (client-only).
• Python 3.10+ required.
• GPU highly recommended for language models.
• Optional extras for data connectors: databricks, googlebigquery, hive, mssql, mysql, oracle, postgres, snowflake.
• Docker support available for isolated environments.
• Documentation: Usage Examples

Highlighted Details

• Broad data support: mixed-type, single/multi-table, time-series, geospatial, text.
• Advanced training: GPU/CPU, differential privacy, progress monitoring, automated QA reports.
• Flexible sampling: up-sampling, conditional generation, re-balancing, rule-adherence.
• TabularARGN models offer 1-2 orders of magnitude efficiency improvement.

Maintenance & Community

• Developed by MOSTLY AI.
• Citation available via BibTeX.

Licensing & Compatibility

• Fully permissive open-source license.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

• Installing local extras (e.g., mostlyai[local]) may downgrade NumPy to 1.26 due to Opacus dependency.
• A runtime restart is needed after installing local extras in environments like Google Colab.