Kumo-Tabular
- 222
- 300k
- openmdw-1.1
- 5 wk ago
A tabular foundation model for corporate credit risk, pretrained only on synthetic data. Every number is published with how it was measured, including where it loses.
Can a tabular foundation model replace your training pipeline? Hands-on TabFM (Google) vs CatBoost vs GPT-6 Astra
Kaggle Playground S6E11 (credit card fraud, ROC-AUC), developed in the open: shared 5-fold CV, in-fold target encoding of exact values and digit keys, shallow trees, GLM and GLM boost, nested-CV stacking, and a FinTFM starter notebook
Benchmarking, optimising and serving the TabFM tabular foundation model online: exact KV cache, torch.compile, FastAPI, Triton, KServe on kind and GKE
Tabular foundation model (TabFM) demo, TabPFN alt.
Runs in the browser on Hugging Face
Raw, zero-shot TabPFN-3.5 vs a 2013-style engineered LightGBM on Kaggle's Blue Book for Bulldozers. Pre-registered, five falsifiable hypotheses.
Fine-tuning TabFM, Google's 1.64B tabular foundation model, on an Apple M4 laptop and on an NVIDIA L4 in GKE: zero-shot vs fine-tuned, laptop vs GPU