PluRel

Spec sheet

Purpose

PluRel is an open-source library for synthesizing relational and tabular data.Read 8 Oct 2026 · github.com

License

The GitHub repository identifies the project license as MIT.Read 8 Oct 2026 · github.com

Compatibility

Generated datasets can be made compatible with RelBench.Read 4 Oct 2026 · star-project.stanford.edu

Configurable output

Its configuration controls table layouts, table and row counts, column counts, and structural causal model parameters.Read 8 Oct 2026 · star-project.stanford.edu

Data types and patterns

Feature generation supports numeric, categorical, and boolean values, with temporal trends, cycles, and fluctuations.Read 4 Oct 2026 · star-project.stanford.edu

Feature types

Generated feature columns support numeric, categorical, and boolean types.Read 8 Oct 2026 · star-project.stanford.edu

Generation stages

It models database schemas with directed graphs, inter-table primary–foreign key links with bipartite graphs, and feature distributions with conditional causal mechanisms.Read 8 Oct 2026 · star-project.stanford.edu

Installation

The library is installed with pip install plurel and requires Python 3.12 or later.Read 8 Oct 2026 · github.com

Installation requirement

The library is installed with pip and requires Python 3.12 or later.Read 4 Oct 2026 · github.com

Intended context

The project targets relational foundation model research and data-driven work with complex multi-table databases.Read 4 Oct 2026 · star-project.stanford.edu

Intended use

The project describes PluRel as a framework for research on relational foundation models and synthetic relational database generation.Read 8 Oct 2026 · arxiv.org

Large-scale generation

A multiprocessing script can generate databases in parallel, with the number of databases and processes configurable.Read 4 Oct 2026 · github.com

Parallel generation

A multiprocessing-based script can generate databases in parallel, with the process count configurable.Read 8 Oct 2026 · github.com

RelBench compatibility

The project says its generated dataset objects are compatible with RelBench.Read 8 Oct 2026 · github.com

Research results

The project reports that synthetic database diversity and pretraining token count both show power-law scaling in relational foundation model pretraining loss.Read 8 Oct 2026 · star-project.stanford.edu

Schema input

It can generate data from an existing SQL schema using the optional schema_file setting.Read 4 Oct 2026 · star-project.stanford.edu

SQL schemas

The configuration can optionally use an existing SQL schema as input.Read 8 Oct 2026 · star-project.stanford.edu

Temporal patterns

Feature generation can model temporal correlations using trend, cycle, and fluctuation components.Read 8 Oct 2026 · star-project.stanford.edu

Transfer limitation

The project states that synthetic pretraining alone is insufficient for robust zero-shot transfer and that continued pretraining on real data is critical for distribution alignment.Read 8 Oct 2026 · star-project.stanford.edu

Where Specifiction places it

• AI Synthetic Data Generators#21 of 256.9