MLJAR AutoML · Machine learning for tabular data
Build better models. Understand every result.
Ask the AI agent in MLJAR Studio to train a model, or start with a few lines of Python. AutoML compares models and prepares reports so you can inspect the results and decide what to use.
- Automated training, tuning, and model comparison
- Reports, explanations, and reusable experiments
- From trained model to a shareable prediction app
Open-source Python package · MIT license · GitHub
AutoML is already in Studio. Just ask.
Load a dataset and tell the AI agent what you want to predict. It can prepare the AutoML code, run training, and open the results for you.
“Train an AutoML model from the loaded dataset.”
Review the target and setup, then explore the models together. The Python code stays visible and editable in your notebook.
Prefer writing code yourself? Choose the Python notebook example instead.
Explore MLJAR Studio →
Explore what your models learned.
Every experiment comes with documentation. Compare candidate models, review validation results, and inspect explanations without assembling a report by hand.
Available diagnostics depend on the task and explanation settings.
Compare models
See validation scores and training times together. Example from our employee attrition tutorial.
Choose how you want to train.
Start with understanding, build a reusable pipeline, or spend more time searching. Four modes give you a practical starting point.
from supervised import AutoML
automl = AutoML(mode="Explain")
automl.fit(X_train, y_train)Modes set defaults. You can configure the algorithms, validation, metric, time budget, and explanation level for your project.
Automate the repetitive work.
AutoML handles the modeling steps while you focus on the question, the data, and how the predictions will be used. The selected mode and configuration determine which stages run.
- 1. Prepare data
- 2. Train and tune
- 3. Compare and explain
- 4. Reuse and share
Preprocessing
Handle missing values and categorical variables, with transformations matched to the data and algorithm.
Model selection
Compare baselines, linear models, trees, boosting, neural networks, and other supported algorithms.
Hyperparameter tuning
Explore model settings and refine promising candidates within your chosen training configuration.
Feature engineering
Generate useful numerical relationships with Golden Features, plus supported text and datetime transformations.
Feature selection
Identify potentially uninformative variables using model-based importance and a random reference feature.
Ensembles and saved models
Combine candidates when enabled, and keep trained models, validation outputs, and reports for reuse.
Inspect performance across groups.
Provide sensitive attributes to evaluate group differences alongside predictive performance. Set a fairness metric and threshold, then inspect the tradeoffs in the report.
Fairness-aware training can try mitigation through sample reweighting. If no model reaches the requested threshold, the report makes that visible. Meeting a metric is one part of a broader review of the model and its intended use.
See the fairness example →
Reports for people and AI workflows.
Alongside visual reports, generate a structured summary in Markdown and JSON. Read it in your notebook, process it in Python, or let an AI agent help you review the experiment.
Model comparisons and feature importance in a readable summary, directly in your notebook.
report = automl.report_structured()
print(report)
Turn your model into an app. Share it in one click.
Generate a prediction app with Mercury, then publish from Studio. After setting up your account and selecting the app files, click Publish to deploy through MLJAR Platform.
Your team can make predictions in a browser. Offer a form for one sample or upload a CSV for batch predictions.
Single prediction
Enter feature values and inspect the prediction with its context.
# Generate the prediction app
automl.app()
# Preview locally, or publish
automl.local_app()
# automl.publish_app()Prefer your own server? Self-host with Docker.
A few lines of Python. A complete experiment.
Use the same open-source AutoML engine in your own scripts and notebooks. This example trains on part of a built-in dataset and predicts on a held-out test set.
pip install mljar-supervisedfrom sklearn.datasets import load_breast_cancer
from sklearn.model_selection import train_test_split
from supervised import AutoML
X, y = load_breast_cancer(return_X_y=True, as_frame=True)
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.25, stratify=y, random_state=42
)
automl = AutoML(mode="Explain", results_path="AutoML")
automl.fit(X_train, y_train)
predictions = automl.predict(X_test)Built for work you can inspect and reproduce.
MLJAR AutoML brings model training, documentation, explanations, and saved artifacts into one workflow. Explore practical examples and the methods behind the package.
Learn AutoML step by step
Work through a complete Python example, from data to models and reports.
Start the tutorial →Explore scientific applications
Discover research using machine learning across healthcare, engineering, and other fields.
Browse research →Inspect the implementation
Explore the MIT-licensed source, training modes, fairness interface, and reporting tools.
View the source →Common questions
What data does MLJAR AutoML work with?
MLJAR AutoML is designed for supervised learning on tabular data: binary classification, multiclass classification, and regression. You provide input features and a target column.
Do I need to write Python?
You can ask the AI agent in MLJAR Studio to prepare and run an AutoML workflow. The generated code remains available to inspect and edit. You can also use the mljar-supervised package directly in Python scripts and notebooks.
Which mode should I start with?
Explain is a useful starting point for exploring your data and understanding model behavior. Perform focuses on practical reusable pipelines, Compete explores a broader search, and Optuna supports intensive hyperparameter tuning.
Are explanations available for every run?
The reports and explanations depend on your mode, task, algorithms, and explanation settings. You can configure explain_level to control additional explanations such as feature importance and SHAP.
Does fairness-aware training guarantee a fair model?
No. It evaluates models against your chosen metric and threshold and can try mitigation through sample reweighting. If no model reaches the threshold, it reports that result. These diagnostics support human review.
Can I save and reuse my models?
Yes. AutoML saves the experiment artifacts, including trained models and reports, in the results directory. You can reopen an existing experiment and use it to make predictions.
Can I share a model as a web app?
Yes. Generate a Mercury prediction app with automl.app(), preview it locally, and publish it through MLJAR Platform. Studio also provides a publishing workflow. You can self-host Mercury apps with Docker.
Is MLJAR AutoML free?
The mljar-supervised Python package is open source under the MIT license. MLJAR Studio and managed app hosting are separate products with their own plans.
Start with your data. Finish with a model you understand.
Let the AI agent in Studio help you get started, or build directly with the Python package.