Mitra-v2 tabular foundation model achieves SOTA performance with smaller size

Researchers have introduced Mitra-v2, a new tabular foundation model that achieves state-of-the-art performance on classification and regression tasks. Trained exclusively on synthetic data, Mitra-v2 utilizes a compact 2D Transformer architecture, enabling it to handle longer contexts and larger feature spaces efficiently. Despite its smaller size compared to industry-scale models like TabFM, Mitra-v2 demonstrates competitive or superior results on benchmarks such as TabArena and TALENT, making it a powerful and broadly applicable open-source option for tabular data problems. AI

IMPACT This release offers a powerful, efficient, and open-source tabular foundation model, potentially accelerating adoption in various real-world classification and regression applications.

RANK_REASON The cluster contains a technical report detailing a new model release and benchmark results. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• Apache Software License 2.0
• EXAONE Tabular
• Hugging Face
• Mitra-v1
• Mitra-v2
• TabArena
• TabFM
• TabICLv2
• TabPFN-3

• model release
• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →