TabPFN-Tydra-4HT

Built with TabPFN.

TabPFN-Tydra-4HT is a hybrid Transformer–State Space Model for in-context tabular classification. It alternates four Hydra and four Transformer layers.

Model Details

• Architecture: [Hydra, Transformer] × 4
• Parameters: 15,999,700
• Representation size: 512
• Maximum features: 100
• Maximum classes: 10
• Training data: fully synthetic
• Fine-tuning on the target dataset: not required
• Paper: Tydra: An Efficient Hybrid Model for Tabular Data

Usage

A tested Tydra inference package and scikit-learn-compatible wrapper are currently under development.

This repository currently provides the model weights and configuration for archival and reproducibility. Tydra uses a custom PyTorch architecture and cannot yet be loaded directly using transformers.pipeline().

Public inference instructions will be added after the wrapper has been released and validated in a clean environment.