Low-code machine learning for Python.
Set up an experiment, compare a dozen models, tune the winner, deploy a pipeline — in under twenty lines. Open-source. No vendor lock-in. Runs on your laptop or your cluster.
pip install pycaretUsed by data teams shipping production models · 8M+ PyPI downloads
The whole AutoML loop in your notebook.
Setup, comparison, tuning, deployment — every step exposes a real sklearn-compatible Pipeline. No wrappers, no magic, no "experiment.predict_model_2()".
- Native sklearn 1.7+ pipelines (no internal wrappers).
- Same OOP API for classification, regression, clustering, anomaly, and time-series.
- Optional dashboard for teams who want a UI on top.
from pycaret.classification import ClassificationExperiment
from pycaret.datasets import get_data
# Sample dataset.
data = get_data("juice")
# Set up an experiment + compare 12 models.
exp = ClassificationExperiment(target="Purchase").fit(data)
top = exp.compare_models(n_select=3)
# Tune the winner. Pipeline-in / Pipeline-out.
tuned = exp.tune_model(top.best, n_iter=20)
# Deploy.
pipeline = exp.finalize_model(tuned.pipeline).pipeline
exp.save_model(pipeline, "production-juice-model")Modern internals, the same simple API.
We rewrote the engine from the ground up against sklearn 1.7 and Python 3.11+. The notebook code you wrote in 3.x still works — it’s just faster and lighter underneath.
Five tasks, one API
Classification, regression, clustering, anomaly, time-series — same Experiment object, same verb names, same pipeline contract.
sklearn-native
Every fitted output is a real sklearn or sktime Pipeline. joblib.dump it. Mount it behind a FastAPI route. Zero wrapping.
Plotly-native diagnostics
Confusion matrix, ROC, residuals, SHAP, decomposition, drift — all interactive, all framework-agnostic figures.
Production-ready
A built-in registry + serving layer turns any fitted pipeline into an HTTP endpoint with rolling p50/p95 stats.
Workspace-aware
Multi-tenant projects, runs, deployments, and audit logs — wired into a FastAPI backend you can self-host.
Lean dependencies
PyCaret 4.0 dropped 50+ legacy dependencies. The core engine installs in seconds.
A cockpit for the rest of your team.
Give analysts and product owners a real UI: experiment design, run comparison, model cards with diagnostic plots, drift monitoring, prediction explorer. Self-hosted. MIT-licensed.
- XGBoost0.943
- LightGBM0.939
- Random Forest0.928
- Logistic Regression0.881
Ready when you are.
Install in seconds. Migrate from 3.x in minutes. Read the docs, browse the API, or jump straight into a notebook.