PriorLabs/tabpfn-extensions

What it solves

It provides a suite of experimental tools to expand the capabilities of TabPFN, a model for tabular data. It addresses limitations in the core library by adding support for complex classification tasks, model interpretability, and advanced statistical analysis.

How it works

These extensions act as a layer on top of either the local PyTorch TabPFN package or the cloud-based TabPFN Client. Depending on the specific extension, it leverages TabPFN's internal mechanisms—such as dense sample embeddings—or uses it as a surrogate model for other processes like Bayesian optimization.

Who it’s for

Data scientists and ML engineers working with tabular data who need more than basic classification or regression, specifically those requiring model explainability, data augmentation, or survival analysis.

Highlights

• Interpretability: Uses SHAP values and feature selection to explain predictions.
• Many-Class Support: Enables classification for problems with more classes than the standard TabPFN checkpoint supports.
• Unsupervised Learning: Tools for outlier detection and data generation.
• Bayesian Optimization: Uses TabPFN as a surrogate model with differentiable Expected Improvement.
• TabEBM: Performs data augmentation using TabPFN-based Energy-Based Models.
• Statistical Testing: Provides p-values for feature relevance testing via pval_crt.