Blog
-
Pocket Data Science: Full Mobile ML & Kaggle Roadmap
Yes, run Kaggle ML on phones. Uncalibrated mobile models hit OOM and rank poorly without proper baselines. Follow our complete…
-
Pocket Data Science X: Mobile GBDT for LLM Preferences
Yes, GBDTs beat raw heuristics. Overlooking lexical complexity and prompt coverage inflates multi-class log loss. See our 58-feature mobile GBDT…
-
Multi-Class Log Loss Explained: Math, MLE & Python Code
Yes, lower log loss is better. Confident wrong probabilities trigger explosive penalties. See the MLE derivation, numerical underflow fix &…
-
Pocket Data Science IX: Kaggle LLM Preference Baselines
Yes, heuristics beat random priors. Length bias inflates log loss if uncalibrated on ties. See our 4-level baseline ladder &…
-
Pocket Data Science VIII: Kaggle Store Sales Quad-Ensemble
Yes, quad-ensembles hit Top 100 on phone. Raw gradient boosting ignores store calendar priors and inflates RMSLE. See the 4-model…
-
Pocket Data Science VII: Store Sales Gradient Boosting on Android Termux
Yes, hybrid residuals beat raw trees. Fitting GBDTs directly on raw store sales creates severe holiday overshoot. See the Ridge-CatBoost…
