Zero-shot models vs XGBoost and LightGBM: benchmark across three UCI datasets

A developer benchmarked two pretrained zero-shot models, Mitra v2 and TabPFN-3.5-Fast, against XGBoost and LightGBM on three public UCI datasets. The tasks covered binary classification of shopping sessions, regression of hourly bike rentals, and multiclass classification of bean varieties. All models used the same 60/20/20 train, validation, and test split with a fixed seed, and each foundation model was tested with raw and engineered features. The classical models were given only a small hyperparameter search, which the author acknowledges could be improved with broader tuning. Results are reported as primary metrics per dataset, with the foundation models scoring close to or above the tree-based baselines in several cases.