Skip to content

catboost_utils

A UX wrapper over CatBoost. Not a fork, not a replacement — a thin layer that fixes the rough edges:

  • Readable errors — C++ traces translated to plain English with actionable hints.
  • Pre-flight validation — catch NaN-in-cat-features, inf, single-class targets before training crashes.
  • Ergonomic custom losses — numpy in, numpy out; numba-jit'ed for all tasks.
  • sklearn pipeline compat — CBXClassifier / CBXRegressor work in Pipeline, GridSearchCV, clone.
  • Structured logging — training output as logger.info records with iteration, learn_loss, test_loss extras.
  • Lossless save/load — best_iteration and metadata survive a round-trip.
  • Exception-safe callbacks — user exceptions surface after fit() instead of being silently swallowed.

Everything is opt-in. Use what helps; mix freely with stock catboost.