This repository contains the code for the ShapPFN model described in Real-Time Explanations for Tabular Foundation Models. At a high level, ShapPFN combines:

• PFN-style tabular transformers (TabPFN / nanoTabPFN-like alternating attention over rows and features)
• An explicit additive decomposition (base + per-feature contributions) via separate decoder heads
• ViaSHAP-style Shapley value regression training, so explanations are part of the model output (not a post-hoc procedure)

As reported in the paper, this approach can match KernelSHAP closely (e.g. R²≈0.96 and cosine≈0.99 on their benchmarks) while being orders of magnitude faster (≈1000×), mantaining the performance of other .

A trained model checkpoint for quick testing is also available at huggingface

pip install -e .

This creates a local dataset under data/ that src/train.py can load via --prior_dir.

bash scripts/generate_data.sh

bash scripts/train_shappfn.sh

The script saves checkpoints under checkpoints/ (see the --out-dir in scripts/train_shappfn.sh).

Runs an OpenML suite evaluation and writes a CSV to outputs/:

bash scripts/eval_openml.sh

Compares model-internal attributions against SHAP’s KernelExplainer on a set of OpenML datasets:

bash scripts/eval_shap.sh