tabpfn

TabPFN is a tabular foundation model designed to perform classification and regression tasks on tabular datasets in a single forward pass. Built on a transformer architecture trained on synthetic datasets, it eliminates the need for complex data preprocessing, feature scaling, or one-hot encoding. The library provides scikit-learn-compatible estimators (`TabPFNClassifier` and `TabPFNRegressor`) that run locally with PyTorch and CUDA support, making it highly optimized for GPU acceleration. It supports datasets with up to 1,000,000 rows and 20,000 features (in version 3.5), handles missing values natively, and offers extensions for interpretability (SHAP-based explanations), unsupervised outlier detection, and learned embeddings extraction.