REaLTabFormer

Overview

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

That flexibility comes with practical requirements: installation and use happen locally, and relational generation depends on matching join keys in parent and child tables. The project asks users to cite its research paper. For other options in this category, see AI Synthetic Data Generators.

Key features

Independent and relational data

For tables with independent observations, REaLTabFormer uses GPT-2. Training stops when the synthetic distribution is close to the real one, giving the process a data-based stopping criterion. That is a useful fit for users generating flat tables, though the framework’s stated remit is tabular data rather than unstructured data.

For relational datasets, it uses a sequence-to-sequence model. This makes it relevant when the relationships across tables need to be represented, rather than treating every dataset as a standalone flat table. The trade-off is a concrete input dependency: parent and child tables must have matching join-key columns.

Validation and privacy-oriented design

Observation validators can filter invalid samples; the project includes a GeoValidator example. This gives users a way to apply checks to generated observations, including geographic validity, rather than relying on synthesis alone. The paper describes target masking to prevent data copying and the Qδ statistic with statistical bootstrapping to detect overfitting. Those are privacy-oriented design measures, not a reason to assume generated data carries no privacy risk.

Local workflow

The package installs from PyPI with pip install realtabformer, and examples use pandas DataFrames as model input. The documented workflow fits a model, saves it locally, then samples synthetic data. This suits users who want a Python-based local process; it is not a hosted, browser-first service.

Pricing

REaLTabFormer costs nothing under the REaLTabFormer plan. It is MIT-licensed software, with no paid tier or free-trial term stated. There are no published seat or generation quotas in the plan terms, so teams should not infer managed-service features from the free price: deployment is self-hosted.

There is a version detail worth checking before installation: the plan describes Python >= 3.7, while the current PyPI package requires Python 3.8 or newer. Users should meet the package requirement. PyPI lists version 0.2.4 as released on January 4, 2026.

Platforms

The package is classified as operating-system independent and the supported platforms include Linux, macOS, Windows, and self-hosted deployment. In practice, the Python requirement and local installation shape the experience more than a dedicated graphical interface would.

Who it's for

REaLTabFormer is a strong fit for researchers and project teams who can prepare pandas DataFrames, run Python software locally, and need either independent tabular synthesis or relational generation. It is less suitable for users seeking a managed web workflow or those whose inputs lack the matching join keys required for relational generation. The project acknowledges World Bank-UNHCR Joint Data Center on Forced Displacement funding for work involving responsible microdata access and synthetic population research.

Pros and cons

• Pros: Separate models address both independent tables and relational datasets, widening its usefulness beyond flat-table generation alone.
• Pros: Observation validators, including a GeoValidator example, let users filter invalid synthetic samples.
• Pros: MIT licensing and a free plan make the framework available without a software charge.
• Cons: Python installation and a local model workflow require technical setup rather than a hosted point-and-click experience.
• Cons: Relational generation depends on matching parent and child join-key columns, limiting use with data that does not meet that structure.
• Cons: Users are asked to cite the research paper when using the project, a consideration for research and project workflows.

Alternatives

Pick Synth Studio if an API or web option matters alongside self-hosting; its free plan includes up to 1M rows per generation. Choose MOSTLY AI if you want a freemium alternative with API and web platforms as well as self-hosting; its free plan has no further limits stated.

NVIDIA ShadowPlay is another free option, for Windows 10 or 11 with a supported GPU and GeForce 551.52 driver or later. SimpleTest may suit users focused on test runs: its free Starter plan allows 50 parallel runs and 500 execution minutes per month, with AI-generated datasets limited to one per test. SynthCity is a free open-source Python library for Linux, installable from PyPI or source.

For a broader hosted plan with stated usage allowances, Synthehol Dataset has a free tier with 1,000 credits and 50K rows per month, among other caps. Tonic Fabricate offers a free plan with $5 monthly credits, basic exports, Discord support, and Tonic Cloud. Synthetic Data Vault is worth considering for a local/API workflow: its free Community plan supports five data types, nine models, CSV and Excel, and forum support.

Verdict

Choose REaLTabFormer if you need a no-cost, locally run Python framework that can synthesize both independent tabular data and relational datasets, and you can meet its input and setup requirements. Its validation tools and two-model approach are the strongest reasons to choose it. Look elsewhere if you need a hosted workflow, or if your relational tables cannot provide matching join keys.