tabular-masterylisted
Install: claude install-skill TranBaDat2607/claude-code-kaggle-grandmaster
# Tabular Mastery
Gradient-boosted trees + great features + diverse ensemble is still the winning recipe.
Neural nets add diversity at the end; on some problems (many numeric features, synthetic
Playground data) a well-tuned MLP/TabM is competitive and strongly boosts the blend.
## 1. Baseline in the first hour
- LightGBM with sensible defaults on raw features (categoricals as `category` dtype),
frozen folds, early stopping, logged to the ledger, one submission.
Template: `templates/train_gbdt.py` (handles LGBM/XGB/CatBoost, OOF, test preds, ledger). It early-stops on a
10% slice of each fold's training part, never the scored fold (`--refit` then refits on all of it), and prints
the per-fold train score next to the CV: the train/validation gap.
- Then CatBoost and XGBoost on the same features — they diverge in categorical handling and
tree growth, giving free ensemble diversity.
- **Diverse baselines on day 1, not at the end**: the same template runs `--model linear`, `knn`,
`svm` and `mlp` (scaled numerics, one-hot categoricals) on the same folds. This shows which families
fit the data. A linear model close to the GBDT suggests a simple signal or a leak, and a large gap
suggests interactions worth engineering. Their OOFs are also blend material: a GBDT + NN + SVR ensemble
with no feature engineering has placed 2nd in a Playground. Re-run them after big data changes as a
leak check.
- Strong default-settings references for the blend: TabPFN (v2.5+ handl