PriorLabs/tabpfn-extensions

TabPFN Extensions – quick overview

What it is – A collection of optional add‑ons that sit on top of the TabPFN library (a fast, transformer‑style tabular‑data predictor). The extensions turn TabPFN into a more general‑purpose toolbox: interpretability, handling many classes, unsupervised tasks, image columns, Bayesian optimisation, etc.

Why it matters – TabPFN already gives you a ready‑to‑use classifier/regressor that works out‑of‑the‑box on tabular data. These extensions let you:

• Explain predictions with SHAP and partial‑dependence plots.
• Push TabPFN beyond its native class‑limit.
• Generate synthetic rows, detect outliers, or impute missing values.
• Feed pictures (as file paths, base64, PIL images…) into a tabular model by automatically converting them to frozen DINOv3 embeddings.
• Use the model’s predictive distribution for Bayesian optimisation or statistical testing.

All of this is experimental (the README warns that APIs may change), but the code is open‑source, pip‑installable, and comes with notebooks and per‑extension READMEs.

---

Core extensions (as listed in the README)

Extension	What it adds
interpretability	SHAP values, feature‑selection utilities, partial‑dependence plots
many_class	Strategies to work with more target classes than the TabPFN checkpoint supports
unsupervised	Data generation, imputation, outlier detection
embedding	Access to TabPFN’s internal dense sample embeddings (local‑only)
image	Automatic conversion of image columns to frozen DINOv3 features before TabPFN sees them
tabebm	Data augmentation via a TabPFN‑based Energy‑Based Model
pval_crt	Conditional randomisation test delivering p‑values for feature relevance
bayesian_optimization	Uses TabPFN as a surrogate model with exact Expected Improvement (local‑only)

---

Getting started

# Install directly from the repo (includes optional deps for all extensions)
pip install "tabpfn-extensions[all] @ git+https://github.com/PriorLabs/tabpfn-extensions.git"

You also need the underlying TabPFN implementation:

• Local inference – pip install tabpfn (full PyTorch version) – required for embedding and bayesian_optimization.
• Cloud inference – pip install "tabpfn-client>=0.2.7" – works for most other extensions.

A Colab notebook is linked in the README and gives a hands‑on walkthrough of installation, classification, regression, and several extensions.

---

Where to look for more detail

• Examples folder – ready‑to‑run scripts/notebooks for each extension (e.g., examples/bayesian_optimization/).
• Per‑extension READMEs – deeper docs live alongside the source code (src/tabpfn_extensions/.../README.md).
• Online docs – selected extensions have full guides at https://docs.priorlabs.ai (e.g., the many‑class guide).
• Community – Discord badge, Twitter, and a contributors list; the project encourages pull‑requests and uses the uv tool for development.

---

Typical use cases

1. Explain a TabPFN model – generate SHAP values and rank features without writing custom code.
2. Work with >100 classes – split or map the problem so TabPFN can still be used.
3. Add image data to a tabular dataset – drop a column of image paths and let the extension turn them into DINOv3 embeddings automatically.
4. Run Bayesian optimisation – treat TabPFN’s predictive distribution as a surrogate and optimise hyper‑parameters or experimental conditions.
5. Synthetic data & outlier detection – generate realistic rows or flag anomalous entries in an unsupervised fashion.

---

License & governance

• License: Apache 2.0 (per the badge and LICENSE file).
• Telemetry: The repo references a telemetry policy, indicating some usage data may be collected (details in a separate TELEMETRY.md).
• Contributions: Welcomed; the project uses uv for environment management and provides a CONTRIBUTING guide.

---

TL;DR

TabPFN Extensions is a genuine, open‑source add‑on suite for the TabPFN tabular‑ML model, offering interpretability, multi‑class handling, image support, unsupervised utilities, Bayesian optimisation, and more. Install via pip, pick the backend you need (local PyTorch or cloud client), and follow the provided notebooks or per‑extension READMEs to start extending TabPFN’s capabilities.