TabPFN-Rel

TabPFN-Rel applies TabPFN to prediction tasks over relational databases. It follows relationships between tables, builds features with Deep Feature Synthesis, and uses TabPFN to predict outcomes for entities such as customers or sellers.

Documentation · Olist cookbook · Technical report

Try the Olist notebook

The task predicts whether a seller active in the past 30 days will receive no orders in the next 30 days. You can also read the walkthrough in the docs or run the local script.

Installation

Requires Python 3.11 or 3.12. Choose a backend:

pip install "tabpfn-rel[api]"    # Hosted inference
# Or: pip install "tabpfn-rel[local]"  # Local inference; GPU recommended

For hosted inference, authenticate before fitting:

from tabpfn_client import init

init()

For local inference, follow the model access guide.

Run the example as a local script

Prefer a Python script to a notebook? The Olist example includes the same database schema and prediction task as YAML files, and reports held-out ROC-AUC. It runs on your machine with either hosted or local inference.

git clone https://github.com/PriorLabs/tabpfn-rel.git
cd tabpfn-rel
uvx kaggle datasets download -d olistbr/brazilian-ecommerce -p data/olist --unzip
uv sync --extra api --group cpu
uv run --no-sync python -c "from tabpfn_client import init; init()"
OMP_NUM_THREADS=1 uv run --no-sync python examples/olist_seller_churn.py

Predict on your own database

The Relational Predictive Interface (RPI) makes TabPFN-Rel easier to use on your own database: define a prediction task, then call fit and predict. RPI is provided by relarena-core, which is installed automatically with TabPFN-Rel.

Define your tables, keys and timestamps in a database YAML file, then define the target and temporal splits in a task YAML file. The task-definition guide describes these formats; the Olist example provides complete files to adapt.

from tabpfn_rel import PredictiveContext, PredictiveQuery, TabPFNRel

context = PredictiveContext.from_yaml("task.yaml", data_dir="data/")
model = TabPFNRel(model="client")
model.fit(context, n_trials=0)
query = PredictiveQuery(entities="all", at_timestamp="test_timestamp")
predictions = model.predict(query)

Use model="local" for local inference. n_trials=0 (the default) fits the default configuration; a positive budget enables temporal tuning. Pass seed and cache_dir to fit to control tuning randomness and feature caching. predict reuses that cache unless given another cache_dir.

model="client" uses TabPFN 3.5 with simple text processing; model="local" uses TabPFN 3.5 with native TF-IDF. Their explicit recipe selectors are:

Prefix the selector with tabpfn-rel- to use it as a RelArena model ID. To use the previous defaults, explicitly select model="client-2026-08-15" or model="local-2026-08-15".

Both query fields are required. Use at_timestamp="test_timestamp" for the context cutoff or an explicit date for another prediction anchor. The database remains frozen at the context cutoff, including for later anchors, to follow RelArena's fixed-snapshot evaluation protocol and prevent post-cutoff data from entering predictions. See RelArena's temporal-validation protocol.

To benchmark TabPFN-Rel against other methods, see RelArena.

Development

uv sync --locked --group cpu
OMP_NUM_THREADS=1 uv run --no-sync pytest
uv run --no-sync pre-commit run --all-files
uv build

Tests cover feature generation, context selection, temporal tuning and prediction without model downloads or hosted API calls.