plurel 1.0.0

Synthesize diverse multi-tabular relational databases using Structural Causal Models, enabling scaling laws for Relational Foundation Models.

PluRel

Synthetic Data unlocks Scaling Laws for Relational Foundation Models

This repository provides a reference implementation for the paper PluRel: Synthetic Data unlocks Scaling Laws for Relational Foundation Models.

The architecture and training code is an improved version of the original implementation for the ICLR 2026 paper Relational Transformer: Toward Zero-Shot Foundation Models for Relational Data.

Overview

PluRel is a framework for synthesizing diverse multi-tabular relational databases using Structural Causal Models (SCMs). This repository provides:

• Scalable generation of synthetic relational data (from scratch or SQL schemas) compatible with relbench.
• High-performance context sampling via a Rust-based sampler (rustler).
• Pretraining of relational transformers on synthetic data.

Framework Design

Setup

Setup the development and testing environment with pixi.

# setup pixi environment
$ pixi install

# Compile and install the rust sampler
$ cd rustler && pixi run maturin develop --uv --release && cd ..

# Run tests
$ pixi run pytest

# Lint and format code
$ pixi run ruff check .
$ pixi run ruff format .

# Install pre-commit hooks
$ pixi run pre-commit install

# link cache repository
$ mkdir ~/scratch
$ ln -s ~/.cache/relbench ~/scratch/relbench

Synthesize Relational Data from Scratch

• The SyntheticDataset class can be used to create relbench compatible dataset objects.
• It only requires a seed and a Config object that contains database, scm and dag level params for sampling. See example below.

from plurel import SyntheticDataset, Config

# create relbench compatible dataset
dataset = SyntheticDataset(seed=0, config=Config())

# create database which can be cached via relbench APIs
db = dataset.make_db()

Configuration

The Config class controls all aspects of synthetic database generation through three parameter groups:

Parameters	Description
DatabaseParams	Table layout (BarabasiAlbert, ReverseRandomTree, WattsStrogatz), number of tables, row counts, column counts, and timestamp ranges.
SCMParams	SCM graph layouts, column types, MLP initialization, activation functions, noise distributions, and time-series trend/cycle parameters.
DAGParams	DAG-specific parameters like edge dropout, in-degree limits, and rewiring probabilities for different graph types.

from plurel import Config, DatabaseParams, SCMParams

config = Config(
    database_params=DatabaseParams(num_tables_choices=Choices(kind="range", value=[5, 10])),
    schema_file="path/to/schema.sql",  # optional: generate from SQL schema
    cache_dir="~/.cache/relbench",       # optional: cache generated databases
)

Scalable Generation

We also provide a multiprocessing-based script to generate databases in parallel.

$ pixi run python scripts/synthetic_gen.py \
    --seed_offset 0 \
    --num_dbs 1000 \
    --num_proc 16 \
    --preprocess

Argument	Description
--seed_offset	Seed offset for database generation. DBs will be named rel-synthetic-<seed>.
--num_dbs	Number of databases to generate.
--num_proc	Number of parallel processes (default: number of CPU cores).
--preprocess	Run preprocessing and embedding steps. Omit to skip.

[!NOTE] Checkout notebooks in examples/ for synthesizing from SQL schemas

Download Preprocessed Data

The preprocessed synthetic data is available on the Hugging Face Hub at kvignesh1420/plurel.

1. Install the HuggingFace CLI (if not present)

pixi add huggingface_hub

2. Create the destination

mkdir -p ~/scratch/pre

3. Download the repository contents into ~/scratch/pre

pixi run hf download kvignesh1420/plurel \
    --repo-type dataset \
    --local-dir ~/scratch/pre

The preprocessed relbench data is available on the Hugging Face Hub at hvag976/relational-transformer.

pixi run hf download hvag976/relational-transformer \
    --repo-type dataset \
    --local-dir ~/scratch/pre

Download Synthetic Pretrained Checkpoints

The synthetic pretrained model checkpoints are hosted on the Hugging Face Hub at kvignesh1420/relational-transformer-plurel.

$ mkdir -p ~/scratch/rt_hf_ckpts

$ pixi run hf download kvignesh1420/relational-transformer-plurel \
    --repo-type model \
    --local-dir ~/scratch/rt_hf_ckpts

One of the downloaded checkpoints will be listed as:

$ ls ~/scratch/rt_hf_ckpts

# model pretrained on a dataset of size 4B tokens curated from 1024 synthetic RDBs
synthetic-pretrain_rdb_1024_size_4b.pt

Pretraining Experiments

• Baseline (real-world) pretraining on relbench datasets with a randomly initialized relational-transformer (RT) model.

$ pixi run torchrun --standalone --nproc_per_node=1 scripts/baseline_pretrain.py

• Synthetic pretraining on varying number of databases and dataset sizes with a randomly initialized RT model.

$ pixi run torchrun --standalone --nproc_per_node=1 scripts/synthetic_pretrain.py

• Continued pretraining on relbench datasets using the synthetic pretrained models. For faster experimentation, the downloaded models from huggingface (stored in ~/scratch/rt_hf_ckpts) can be passed to the load_ckpt_path argument in the training script.

$ pixi run torchrun --standalone --nproc_per_node=1 scripts/cntd_pretrain.py

Citation

If you find this work useful, please cite our paper:

@article{kothapalli2026plurel,
  title={{PluRel:} Synthetic Data unlocks Scaling Laws for Relational Foundation Models},
  author={Kothapalli, Vignesh and Ranjan, Rishabh and Hudovernik, Valter and Dwivedi, Vijay Prakash and Hoffart, Johannes and Guestrin, Carlos and Leskovec, Jure},
  journal={arXiv preprint arXiv:2602.04029},
  year={2026}
}

If you use the architecture, training loop or sampler code, please also cite the Relational Transformer paper:

@inproceedings{ranjan2026relationaltransformer,
    title={{Relational Transformer:} Toward Zero-Shot Foundation Models for Relational Data}, 
    author={Rishabh Ranjan and Valter Hudovernik and Mark Znidar and Charilaos Kanatsoulis and Roshan Upendra and Mahmoud Mohammadi and Joe Meyer and Tom Palczewski and Carlos Guestrin and Jure Leskovec},
    booktitle={The Fourteenth International Conference on Learning Representations},
    year={2026}
}