Prior Labs, the research division of Hugging Face, has announced its first foray into relational learning with the release of three open-source software components. The suite aims to accelerate research and practical adoption of relational learning, which deals with predictions across interconnected data tables.
The centerpiece is RelArena-α, a unified benchmark framework built on RelBench v1. Inspired by tabular benchmarks like TabArena, it standardizes data loading, evaluation protocols, and tuning regimes, enabling fair and reproducible comparisons of relational learning methods. The framework also supports custom tuning systems.
Alongside the benchmark, Prior Labs introduces TabPFN-Rel, a relational harness for the TabPFN-3 model. Currently ranked first among models on the RelArena-α leaderboard, TabPFN-Rel demonstrates that flattening a relational database into a single table can remain competitive with specialized relational architectures. The leaderboard also lists RT-PluRel as the top system submission.
To ease adoption, the Relational Predictive Interface (RPI) provides a model-agnostic, open-source API. It allows users to define prediction tasks on their own databases and run any RelArena-α model, such as TabPFN-Rel, with just a few lines of code, packaged as a simple PyPI library.
This alpha release targets researchers and early adopters. Prior Labs plans to iterate based on community feedback, aiming to evolve RelArena-α into a catalyst for progress in relational learning.