Mitra-v2: A tabular foundation model leading TabArena and ScoringBench

Details

Mitra-v2 is a second-generation tabular foundation model for classification and regression, released under the Apache-2.0 license on Hugging Face by the research team. Amazon Connect Customer led the effort and is adopting the model for structured business data.
Mitra-v2 leads the single-model field on TabArena, a benchmark of 51 real-world datasets, and ranks first on ScoringBench for regression tasks.
The model outperforms Mitra-v1 on 32 of 33 datasets where both run natively and completes all 51 TabArena datasets within a one-hour budget, a task Mitra-v1 failed on more than a third of.
Pretraining uses synthetic tables with 10 times as many rows and roughly three times as many columns as Mitra-v1, generated via a new hybrid SCM prior that mixes neural networks, decision trees, Gaussian processes, and autoregressive processes.
Deployment includes automatic fine-tuning with bagging (eight averaged copies) and uses singular-value decomposition for wide tables or a hierarchical tree for tasks with more than 10 classes.

Key point