REaLTabFormer

REaLTabFormer is a free, open-source Python framework for generating synthetic tabular data, including related tables. It is best suited to research and project teams comfortable working in Python; its chief advantage is offering separate approaches for independent and relational data in one self-hosted package.

See our AI Synthetic Data Generators list for other options.

Overview

REaLTabFormer covers two data shapes: independent observations in a flat table and records linked across parent and child tables. Its workflow is model-focused: fit a model, save it locally, and sample synthetic data. The package is operating-system independent according to PyPI, while its deployment model is self-hosted rather than a hosted service.

The project describes use in projects and research and asks users to cite its research paper. It acknowledges World Bank–UNHCR Joint Data Center on Forced Displacement funding for work involving responsible microdata access and synthetic population research.

Key features

Independent and relational data

For independent tabular observations, REaLTabFormer uses GPT-2. Training stops when the synthetic distribution is close to the real one, making the stopping behavior distribution-based. That may suit users whose goal is to approximate a particular dataset, but the described criterion does not promise any particular quality threshold or outcome.

For related tables, a sequence-to-sequence model generates relational datasets. This is a meaningful distinction from a flat-table-only workflow, but it depends on matching join-key columns in the parent and child tables. If the source tables do not have corresponding keys, this approach is not a fit without addressing that prerequisite.

Validation and privacy-oriented design

Observation validators can filter invalid synthetic samples; the GeoValidator example is intended to remove invalid generated samples. The paper describes target masking to prevent data copying and use of the Qδ statistic with statistical bootstrapping to detect overfitting. These are privacy-oriented design measures, not a guarantee that generated data is safe for every release or use case.

The package installs from PyPI with pip install realtabformer. Its current PyPI package requires Python 3.8 or newer. The package is distributed under the MIT License, and PyPI lists version 0.2.4 as released on January 4, 2026.

Pricing

REaLTabFormer costs free. The free plan is the open-source package under the MIT License; there is no separate paid tier or free-trial period in the stated model. This suits users able to run and maintain a self-hosted Python package, rather than those seeking a managed service.

The plan description gives a Python >= 3.7 requirement, while the current PyPI package requires Python 3.8 or newer; use the package requirement when preparing an installation. No seat or usage quota is stated for the free plan.

Platforms

REaLTabFormer supports Linux, macOS, and Windows, and is self-hosted. PyPI classifies the package as operating-system independent. The practical requirement is a compatible Python environment, with pandas DataFrames used as model input in the examples.

Who it's for

Choose REaLTabFormer if you need a self-hosted, MIT-licensed tool for synthetic independent tables or relational datasets, and can work with Python and pandas. It is particularly relevant when linked tables matter and you can provide matching join keys. It is a weaker fit if you want a hosted interface or do not control the shape and keys of your input data.

Pros and cons

Pros

Two data-generation approaches: GPT-2-based modeling covers independent observations, while a sequence-to-sequence model addresses relational datasets.

Free and MIT-licensed: there is no listed paid plan or free-plan quota to weigh against the open-source package.

Validation support: observation validators, including a GeoValidator example, can filter invalid synthetic samples.

Privacy-oriented methods: the paper describes target masking and a bootstrapped Qδ statistic for detecting overfitting.

Cons

Requires a Python workflow: installation and pandas-based inputs favor technically equipped teams over users seeking a ready-made hosted product.

Relational generation has a key prerequisite: parent and child tables need matching join-key columns.

Python minimums differ by plan and package description: the current package requires Python 3.8 or newer, despite the plan description's Python >= 3.7.

Alternatives

FakerForge is a freemium alternative with a free monthly plan capped at 1 database, 50,000 AI tokens, 100 rows per table, 10 tables per database, 500 API requests, 1 mock API, and 5 endpoints per mock; consider it if those stated limits and its API, web, or desktop platform options suit your workflow better.

Synth Studio is another free, MIT-licensed option, with up to 1M rows per generation and API, web, or self-hosted platforms; consider it if that generation cap and its platform choices match your needs.

SimpleTest has a free Starter plan with 50 parallel test runs and 500 execution minutes per month, while AI-generated datasets are limited to one per test; choose it when those test-run limits suit the job.

Synthehol Dataset offers a free plan with 1,000 credits and 50K rows per month, 5 projects, 5 datasets, 1 user, 1 GB storage, and 30-day retention; its stated CSV and JSON formats may make it a better fit for that bounded workflow.

IRI DarkShield is a paid alternative with a free trial and a perpetual-use standalone license whose first-hostname fee ranges from USD $5K-50K; consider it if that licensing model fits.

NVIDIA ShadowPlay is a free Windows-only alternative requiring Windows 10 or 11, GeForce 551.52 driver or later, and a supported GPU.

Anyverse is a paid, web-based alternative with a free trial.

MOSTLY AI has a free plan and supports API, Linux, self-hosted, and web use; consider it if those platform options suit you better.

Verdict

REaLTabFormer is a strong choice for research and project teams that can manage a Python installation and want free, self-hosted synthesis for either independent tables or linked datasets. Its breadth and validation tools are compelling, but the Python workflow and join-key requirement are real constraints. Look elsewhere if you need a hosted product or cannot meet those input requirements.