Why your fine-tuning job takes longer than the FLOPs say
This article explains why fine-tuning jobs take longer than FLOP-based estimates predict, presenting a CatBoost model that achieves 98% R-squared accuracy by incorporating evaluation and checkpoint overhead features.