TabTune: A Unified Library for Inference and Fine-Tuning Tabular Foundation Models

  • 2025-11-05 17:36:30
  • Aditya Tanna, Pratinav Seth, Mohamed Bouadi, Utsav Avaiya, Vinay Kumar Sankarapu
  • 0

Abstract

Tabular foundation models represent a growing paradigm in structured datalearning, extending the benefits of large-scale pretraining to tabular domains.However, their adoption remains limited due to heterogeneous preprocessingpipelines, fragmented APIs, inconsistent fine-tuning procedures, and theabsence of standardized evaluation for deployment-oriented metrics such ascalibration and fairness. We present TabTune, a unified library thatstandardizes the complete workflow for tabular foundation models through asingle interface. TabTune provides consistent access to seven state-of-the-artmodels supporting multiple adaptation strategies, including zero-shotinference, meta-learning, supervised fine-tuning (SFT), and parameter-efficientfine-tuning (PEFT). The framework automates model-aware preprocessing, managesarchitectural heterogeneity internally, and integrates evaluation modules forperformance, calibration, and fairness. Designed for extensibility andreproducibility, TabTune enables consistent benchmarking of adaptationstrategies of tabular foundation models.

 

Quick Read (beta)

loading the full paper ...