FAQ

Does it use the GPU?

No.

How does it compare to CatBoost, LightGBM, and XGBoost?

On defaults: roughly around LightGBM/XGBoost or better, and significantly faster than CatBoost. Setting n_ensembles=8 bags the model — on the Grinsztajn benchmark suite this is the strongest model on every accuracy column, ahead of CatBoost at well under half its fit cost on high-cardinality data. Note that since TabArena uses ensembling, this improvement does not show up on those leaderboard results.

Do I need to one-hot encode categoricals or impute missing values?

No.

Pass your categorical columns to fit(..., cat_features=[...]), by integer position or by column name. NaNs route to a dedicated bin at fit and predict time, so no imputation is needed. pandas nullable dtypes (Int64/Float64/boolean) and their pd.NA are accepted and treated as missing.

How can I make inference faster?

If you have already validated your serving data and want to skip it, use scikit's 'assume_finite'.

import sklearn
with sklearn.config_context(assume_finite=True):
    preds = model.predict(X)        # finiteness scan skipped

Why is the very first fit or predict slow?

That is numba compiling the kernels (one-time per process, disk-cached per user). Call chimeraboost.warmup() — or set CHIMERABOOST_WARMUP=1 — to pay it at import time instead of on the first call. See Deployment for numbers and short-lived-worker patterns.

Why oblivious (symmetric) trees?

They make prediction extremely fast and provide strong built-in regularization, at some cost to per-tree sharpness. See How it works.

Does SHAP support multiclass?

Not yet.

How do I save and load a model?

A fitted estimator pickles like any scikit-learn object:

import joblib
joblib.dump(model, "model.joblib")
model = joblib.load("model.joblib")

What exactly does it depend on?

NumPy, numba, scikit-learn, and SciPy. pandas is not required — DataFrames are consumed through their own conversion methods, so passing them works whenever you have pandas (or polars) installed yourself.

How do I tune it?

Mostly, you don't: the defaults are benchmark-tuned, and in our experiments broad hyperparameter search bought little that generalized. Two settings address specific situations rather than general tuning: n_ensembles=8 is the benchmarked maximum-accuracy mode (at several times the fit cost), and depth=8–10 suits large, interaction-heavy regression. Parameters documents every knob.

Is the API stable?

ChimeraBoost is beta (0.x). Breaking API or behavior changes bump the minor version and are recorded in the CHANGELOG; patch releases are fixes only. Pickled models are not guaranteed to load across versions — store the version next to the model and re-fit after upgrading.