relarena
Copyright 2026 PriorLabs GmbH

Licensed under the Apache License, Version 2.0. See the LICENSE file.

This product includes third-party software. The following components are used
under the MIT license, with the upstream copyright notice and license text
retained alongside the code in this package:

  * RelBench example GNN building blocks
    Copyright (c) 2023 RelBench Team
    https://github.com/snap-stanford/relbench
    src/relarena/models/_shared/gnn/_vendor/gnn.py

  * RelGNN composite-message-passing GNN
    Copyright (c) 2023 RelBench Team
    https://github.com/snap-stanford/RelGNN
    src/relarena/models/relgnn/_vendor/

  * RelGT Relational Graph Transformer
    Copyright (c) 2025 Vijay Prakash Dwivedi, Sri Jaladi, Yangyi Shen,
    Federico López, Charilaos I. Kanatsoulis, Rishi Puri, Matthias Fey,
    Jure Leskovec
    https://github.com/snap-stanford/relgt
    src/relarena/models/relgt/_vendor/

  * RDBLearn (method adaptation; no files copied)
    Copyright (c) HKU Shanghai X-Lab
    https://github.com/HKUSHXLab/rdblearn
    src/relarena/models/rdblearn/model.py, src/relarena/models/_shared/tfm/tfm.py,
    src/relarena/featurization/dfs.py

  * Relational Transformer (RT) (dependency; no files copied)
    Copyright (c) 2025 Stanford STAR / Relational Transformer authors
    https://github.com/rishabh-ranjan/relational-transformer
    src/relarena/models/rt/ -- installed from the `rt` extra, not vendored.

Datasets are not distributed with this package. They are downloaded at runtime by
relbench and remain subject to their own upstream terms.

Vendored component provenance
-----------------------------

RelBench example GNN building blocks were copied from
https://github.com/snap-stanford/relbench at commit 74d4c37
(`examples/model.py` and `examples/text_embedder.py`). Their imports were merged
into `src/relarena/models/_shared/gnn/_vendor/gnn.py`; only Ruff formatting was
otherwise normalized. The GNN layers themselves are imported from the installed
`relbench.modeling.nn` package and are not copied.

RelGNN model building blocks were copied from
https://github.com/snap-stanford/RelGNN at commit cffdb8b. The upstream example
modules were split into the matching modules under
`src/relarena/models/relgnn/_vendor/`; package imports were rewritten and Ruff
formatting normalized. Duplicate RelBench encoders were not copied.

RelGT model and sampler building blocks were copied from
https://github.com/snap-stanford/relgt at commit 19e423ca. Upstream `utils.py`
became `_sampler.py` and retains only the sampler functions and globals RelArena
uses; inter-module imports were rewritten to package paths and Ruff formatting
normalized. `models/relgt/model.py` and `models/relgt/tokenize.py` are original
RelArena adapter code, not vendored files.

RDBLearn is a method adaptation rather than a file copy. RelArena reimplements
the Deep Feature Synthesis to tabular-foundation-model recipe from
https://github.com/HKUSHXLab/rdblearn. The closest adaptation is the
non-stratified train downsampling in `models/_shared/tfm/tfm.py`; RelArena's DFS
cache and TabPFN-Rel extensions are original work.
