RDB-PFN · Relational Database PFN

Predict over linked tables in-context · no per-database training

Use when the signal for a prediction is spread across several linked tables — customers, orders, events, sensor logs — and you want a zero-shot baseline before building a bespoke feature pipeline. Provide the tables, their relationships, and the target (table + id + label column). Inference-only: the upstream repo ships weights and inference code but no training prior, so there is nothing to continue-pretrain — pick TCPFN or Do-PFN for a trainable base model.