What it does imodels collects roughly two dozen interpretable classifiers and regressors under one scikit-learn-compatible roof. The API is deliberately boring: fit, predict, predict_proba, print(model). The payoff is that the fitted models spit out actual rules or tiny trees you can read over coffee.

The interesting bit The package doesn’t just wrap CART. It pulls in research implementations—Bayesian rule lists, optimal sparse trees (GOSDT), hierarchical shrinkage for post-hoc regularization, FIGS (sums of small trees)—that are otherwise scattered across individual repos and half-finished theses. The “distillation” and “AutoML” wrappers are a nice touch: train the black box first, then squeeze it into something auditable.

Key highlights

• Full sklearn compatibility: drop-in replacements for pipelines, GridSearchCV, etc.
• Covers four model families: rule sets, rule lists, rule trees, and algebraic models (e.g., sparse integer linear models)
• Includes recent research methods: FIGS, hierarchical shrinkage, RF+ feature importance
• Post-hoc tools: distillation wrapper, AutoML wrapper, uncertainty estimation demos
• Published in JOSS; supports Python 3.9–3.11

Caveats

• Some Bayesian methods are explicitly marked “slow” in the docs
• A few listed models are “coming soon” (Lightweight Rule Induction, MLRules)
• The README claims accuracy is often not sacrificed, but no benchmarks are shown—your mileage will vary by dataset

Verdict Worth a look if you work in regulated industries, healthcare, or anywhere a regulator asks “why did it say no?” Not for you if your only metric is Kaggle leaderboard position and you don’t care which features matter.