fast-experiment: Kaggle Tabular Toolkit

The core idea

Kaggle tabular work is the same loop every time: load, EDA, feature engineering, baselines, ensembling, pseudo-labeling. fast-experiment collapses that loop into a handful of calls, but the real design decision is the contract underneath it.

Every model emits the same artifact: an OOFResult (out-of-fold predictions plus test predictions). Once everything speaks that one language, stacking, hill climbing, and pseudo-labeling stop caring what produced the numbers. They become plain array math over a shared interface. Add a new estimator and the entire ensembling stack works with it for free.

What’s inside

• data / eda: cuDF-aware loading with memory reduction, plus leakage_report, adversarial_validation, and PSI drift checks.
• features: out-of-fold TargetEncoder (leak-free by construction), count and label encoders.
• estimators / baselines: one GBDTEstimator wrapping XGBoost, LightGBM, and CatBoost with device="auto"; run_baselines sweeps them in a single call.
• ensemble: Caruana greedy hill_climb and a cuML/sklearn stack meta-learner.
• tuning / pseudo: Optuna search that refits into an OOFResult, and an iterative confident-pseudo-label loop.

GPU-first, CPU-safe

RAPIDS (cuDF/cuML/cuPy) is used automatically when a GPU is present and falls back to CPU when it isn’t. The same code runs on a laptop and on a Kaggle GPU kernel. No branching in user code, no separate paths to maintain.

Why I built it

Competition iteration speed is mostly a function of how little friction sits between “I have an idea” and “I have an out-of-fold score for it.” A single artifact contract plus automatic GPU acceleration removes most of that friction. It’s pre-1.0 and tabular-only for now. Deep-learning estimators are planned behind the same OOFResult contract, which is the whole point: the interface was designed to absorb them without touching the ensembling code.