REaLTabFormer

REaLTabFormer is an open-source framework for generating synthetic tabular and relational datasets. For relational data it uses a sequence-to-sequence model; its model for independent tabular observations uses GPT-2. Examples use pandas DataFrames as input, and relational generation needs matching join-key columns in the parent and child tables. The documented workflow fits a model, saves it locally, and samples synthetic data. For non-relational training, the model stops when the synthetic distribution is close to the real distribution. The framework includes observation validators that can filter invalid samples, with a GeoValidator example. The paper describes target masking to prevent data copying and the Qδ statistic with statistical bootstrapping to detect overfitting. The package is free under the MIT License and can be installed from PyPI with pip install realtabformer. The current PyPI package requires Python 3.8 or newer and is classified as operating-system independent; Linux, macOS, Windows, and self-hosted use are listed. The project describes use in projects or research and asks users to cite its research paper when using it.

REaLTabFormer provides distinct approaches for independent tabular observations and relational datasets, plus validation tools. It is a free Python package, with join-key requirements for relational generation.

REaLTabFormer suits researchers or project teams that need synthetic tabular or relational data and can work with Python and pandas. Relational use requires matching join keys between parent and child tables.

Overview

REaLTabFormer is a Python package for generating synthetic tabular data, including datasets split across related tables. It suits researchers and project teams who can work in Python and prepare data as pandas DataFrames. Its main advantage is supporting both independent observations and relational data in one free, self-hosted framework; relational generation requires matching join keys.

The project describes its use in research and projects, and asks users to cite its research paper. It acknowledges World Bank–UNHCR Joint Data Center on Forced Displacement funding for work on responsible microdata access and synthetic population research. Browse our AI Synthetic Data Generators list to compare other tools in the category.

Key features

Models for flat and relational data

For tables of independent observations, REaLTabFormer uses GPT-2. Training stops when the generated distribution is close to the real one, a useful data-driven stopping rule for users who do not want to choose a fixed training duration. For related tables, its sequence-to-sequence model generates relational datasets. That makes the package relevant when a flat table cannot represent the data relationships, but the parent and child tables must have matching join-key columns.

Sampling, validation, and privacy design

The documented workflow fits a model, saves it locally, then samples synthetic data. Observation validators can filter invalid samples, with a GeoValidator example for geographic data. This validation interface helps address constraints that a model's generated distribution alone may not enforce.

The paper describes target masking to prevent data copying and the Qδ statistic with statistical bootstrapping to detect overfitting. These are privacy-oriented design measures, not a guarantee that every generated dataset is safe to share. Teams should weigh them alongside their own data governance needs.

Python package and license

Installation is through PyPI with pip install realtabformer. The current PyPI package requires Python 3.8 or newer and is classified as operating-system independent; the software is self-hosted and distributed under the MIT License. The package plan description says Python 3.7 or newer, so users should follow the current package requirement when installing.

Pricing

REaLTabFormer costs nothing. The free plan provides MIT-licensed software for self-hosted use, with no free trial. It is a strong fit for Python users who want to run synthetic data generation in their own environment without a subscription. Its cost advantage comes with the practical work of installing, preparing inputs, fitting, saving, and sampling models yourself.

Platforms

REaLTabFormer supports Linux, macOS, and Windows, and is self-hosted. Its operating-system-independent PyPI classification and Python package requirements make it suited to teams comfortable with a Python workflow rather than those seeking a hosted interface.

Who it's for

This package is best for researchers and project teams generating synthetic tabular data who can prepare pandas DataFrames and manage Python software locally. It is especially useful when the work needs both independent rows and linked parent-child tables, or when validators are needed to filter domain-invalid samples. It is a poor fit for users seeking a no-code service or relational generation without aligned join keys.

Pros and cons

• Pros: One MIT-licensed package supports GPT-2-based independent tables and sequence-to-sequence relational generation, covering two data structures in a self-hosted workflow.
• Pros: Distribution-based stopping for non-relational training and observation validators offer concrete controls over training and sample validity.
• Pros: The paper describes target masking and a bootstrapped Qδ overfitting check as privacy-oriented design choices.
• Cons: Relational generation depends on matching join-key columns between parent and child tables, adding a prerequisite to data preparation.
• Cons: Installation, model fitting, and sampling require a Python environment and local workflow, which may not suit users who want a hosted product.
• Cons: Python requirements differ between the package plan description (3.7 or newer) and the current PyPI package (3.8 or newer); follow the current package requirement.

Alternatives

Choose DataSynthesizer if you want another free, MIT-licensed package installable with pip and your input is a first-normal-form table. For a hosted option, Synthehol Dataset has a free plan with 1,000 credits, 50K rows per month, five projects, five datasets, one user, 1 GB storage, 30-day retention, and CSV and JSON support.

SynthEngyne may suit users who want a monthly free-credit model: its free plan provides 1,000 credits each month, with text, tabular, and time-series rows costing one credit each and images costing 10 credits each. Synthetic AI Data has a free plan for getting started, with specific limits unstated.

Tabularis.AI is a paid alternative with a free developer plan offering 10,000 credits per month, while Tonic Fabricate has a free plan with $5 monthly credits, basic data exports, Discord support, and Tonic Cloud. SimpleTest is another freemium option, with a free starter plan capped at 50 parallel test runs and 500 execution minutes per month, plus one AI-generated dataset per test. NVIDIA ShadowPlay has a free Windows plan for users meeting its GeForce driver and supported GPU requirements.

Verdict

REaLTabFormer is a compelling free choice for Python-literate research and project teams that need synthetic independent or relational tables and can prepare matching join keys where needed. Its combination of two model approaches, local sampling, and validation tools is the clearest reason to choose it. Look elsewhere if your priority is a hosted workflow, or if relational table preparation and local Python operation are a poor fit.