Skip to main content
Most real-world data does not live in a single table. It lives in a relational database, where the rows you want to predict sit in one table and the evidence sits in others, reachable through foreign keys. TabPFN-Rel α is the harness that applies TabPFN 3.5 to this setting: it flattens the database into one table and predicts in-context, with no task-specific training.

How it works

TabPFN-Rel runs in three stages.
  1. Deep feature synthesis. Starting from the table that holds the entity you are predicting for, it follows the schema’s foreign keys outward and aggregates the rows reachable at each depth into features, such as COUNT(results) or MEAN(results.points). This is the recipe popularized by featuretools, and TabPFN-Rel builds on RDBLearn’s implementation of it.
  2. Depth. How far to traverse is the one hyperparameter that matters. The current recipes aggregate up to four hops from the entity table.
  3. In-context prediction. The resulting flat table goes to TabPFN 3.5, which predicts the labels of the query rows from the labelled context rows in a single forward pass.
Only rows written before each entity’s prediction timestamp are aggregated, so a model never sees the future of the entity it is predicting. For signatures, parameters and return types, see the TabPFN-Rel Python reference.

Getting Started

TabPFN-Rel predicts on your own relational database through the Relational Predictive Interface (RPI). Use Python 3.11 or 3.12. Choose an inference backend:
For hosted inference, authenticate with tabpfn-client before fitting:
Local inference requires access to the model weights; follow the model access guide. Define your tables, keys and timestamps in a database YAML file, then define the prediction target and temporal splits in a task YAML file. The task-definition guide describes the file formats. With those files and your data in place:
Coding agents such as Claude Code or Codex can automate most of this setup. Install the TabPFN-Rel agent skill, point the agent at your data, and it can clean up the tables, write both YAML files, formulate the label query and compare TabPFN-Rel with baselines. Review the result before fitting. TabPFN-Rel assumes static rows: every value in a row was written at the row’s timestamp and not changed afterwards. Drop any column that clearly violates this, such as an order status or delivery date filled in after the order was placed.
RPI is provided by relarena-core, a dependency shared by TabPFN-Rel and RelArena. It is installed automatically. Use TabPFNRel(model="local") for local inference. Both "client" and "local" select the latest TabPFN 3.5 recipe; pass a dated selector such as model="client-2026-09-28" to pin one. The current recipes run one validated configuration, so n_trials=0 is all they need. Hosted fits and predictions consume API quota. The TabPFN-Rel cookbook walks through a complete database example, including task definitions, feature caching and baseline comparisons.

Key Features

  • No task-specific training: features are computed once and the prediction is a single in-context forward pass, so there is no gradient training loop to configure.
  • Schema-driven: the aggregations follow the foreign keys already declared in your database, so no manual feature engineering is required to get a first result.
  • Nothing to tune: the current recipes run one validated configuration, in contrast to the graph neural networks typically applied to this setting.
  • Text columns included: free-text fields in the entity table are passed to the model rather than dropped, which matters on tasks where the text carries the signal.

Learn More

Advancing Open and Reproducible Relational Learning

The report covering RelArena-α, TabPFN-Rel and the RPI.

TabPFN-Rel on GitHub

Model source, installation and usage.