rdb-pfn
RDB-PFN — Relational Database PFN
pretrained relational-DB PFN · predict over your tables in-context
Choose this when the signal for a prediction lives across several related tables joined by foreign keys, and you want a fast zero-shot baseline. Provide the tables, their relationships, and the target (table + id + label column). Inference-only.
Run it
$ pip install pfnstudio
$ pfnstudio run rdb-pfn --input <your file>
# ex: each task on the sample input pfnstudio ships
$ pfnstudio run rdb-pfn --task classify --input sample
$ pfnstudio run rdb-pfn --task classify_relational --input sample
# download it without running · its tasks, inputs, license and digest
$ pfnstudio pull rdb-pfn
$ pfnstudio show rdb-pfn
# models loaded now · delete it from this machine
$ pfnstudio ps
$ pfnstudio rm rdb-pfn
Python 3.10+. The first run downloads the model's adapter, asks before installing PyTorch and the model's own packages (-y installs without asking) and pulls the pretrained weights from Hugging Face. --input sample runs the sample input pfnstudio ships and prints its path: copy it, put your data in, and pass it as <your file>. Later runs reuse the cache.
Serves
| Task | Answers |
|---|---|
| classify | which class is this row? |
| classify_relational | which class, across linked tables? |
Provenance
| Paper | Relational In-Context Learning via Synthetic Pre-training with Structural Prior ↗ |
| Venue | ICML, 2026 |
| Authors | Yanbo Wang, Jiaxuan You, Chuan Shi, Muhan Zhang |
| Publisher | MuLab · Peking University ↗ |
| Weights | pretrained RDB-PFN · model.pt from GitHub |
| Input | relational DB: tables + FK relationships → target entity label |
| Adapter | basemodels/rdbpfn/adapters/ · rdbpfn |
| Source | github.com/MuLabPKU/RDBPFN ↗ |
Readme
A ready-to-use predictor for data spread across many linked tables — customers, orders, events, sensors. Point it at your relational database and it predicts an outcome for each entity without hand-building features or training a bespoke model per database.
RDB-PFN (Wang, You, Shi, Zhang — 'Relational In-Context Learning via Synthetic Pre-training with Structural Prior', ICML 2026) is a prior-fitted network for relational databases, pretrained purely on synthetic relational data (2M+ tasks) under a structural prior over tables and foreign keys. This listing installs the real released weights as an inference-only base model served through the upstream inference classes (RDBPFNClassifier for a single table; RDBPFNRelationalClassifier + Featuretools DFS for multi-table). Listed with the authors' permission. Continued pretraining is not offered — the upstream repo ships inference code + weights but no training prior. The upstream repo is not pip-installable, so the adapter makes it importable via RDBPFN_REPO_DIR or a one-time clone; the predict runner needs torch/pandas/scikit-learn/featuretools. NOTE: the studio inference path wraps the verified upstream API but has not yet been run end-to-end here — validate on a runner before production use (see the base model's README).