AutoTrain MLRepeatable runs
Each experiment keeps its task, method, parameters, and output together.
Automated machine learning
Automate repetitive ML work without hiding the experiment
Automation should remove repetition, not understanding. AutoTrain records the dataset, configuration, metrics, and artifact behind each experiment.
Workflow
A practical model workflow
AutoTrain keeps preparation, training, comparison, testing, and export attached to the same project so the experiment can be reopened from another device without rebuilding the setup.
Comparable metrics
Review task-appropriate metrics and visualizations before selecting a model.
Operational output
Move from a saved experiment to testing, model export, or API prediction.
How it works
1. Define the target
Define the target and task.
2. Select preprocessing and
Select preprocessing and candidate algorithms.
3. Run experiments with
Run experiments with a consistent validation split.
4. Compare results and
Compare results and keep the model that matches the product objective.
Coverage
Supported tabular tasks
Choose the task from the problem definition. AutoTrain does not silently change the method or hyperparameters selected by the user.
| Task | Typical target | Evaluation focus | Testing output |
|---|---|---|---|
| Classification | Category or class | Accuracy, precision, recall, F1, confusion matrix, ROC/PR where available | Predicted class, confidence, probabilities |
| Regression | Continuous numeric value | MAE, RMSE, R², residual behavior | Predicted numeric value |
| Clustering | No supervised target | Silhouette, cluster size, stability, interpretability | Assigned cluster and distance context |
Saved output
What remains available after training
Experiment record
Dataset reference, preprocessing settings, task, algorithm, hyperparameters, validation metrics, and generated visualizations remain attached to the saved experiment.
Testing interface
Use custom single-row values or a schema-matched CSV to test new data. Results can be exported as CSV for downstream review.
Model artifact
Completed models can be exported with the preprocessing and input schema required to reproduce predictions correctly.
Prediction endpoint
Supported models expose a schema endpoint and a prediction endpoint that can be called from an application backend.
Methodological basis
Automation with explicit controls
AutoML research describes automation across data preparation, model selection, optimization, and evaluation. AutoTrain applies that principle selectively: repetitive execution is automated while task, preprocessing, algorithm, and parameter choices remain visible.
He, X., Zhao, K., & Chu, X. (2021). AutoML: A survey of the state-of-the-art. Knowledge-Based Systems, 212, 106622. Read the journal article
Important: AutoTrain helps execute and document experiments. You remain responsible for data rights, leakage checks, metric selection, bias review, and production monitoring.
FAQ
Common questions
Does AutoTrain choose the best model automatically?
No. It reports comparable experiment results; the user chooses methods, parameters, and the final model based on the actual objective.
Can an experiment be reopened on another device?
Yes. Saved datasets, experiments, model metadata, visual results, and testing tabs are synchronized through AutoTrain Engine.
Is a benchmark result a production guarantee?
No. Benchmarks are fixed reference runs. Real performance depends on data quality, the split strategy, leakage, drift, and the cost of each error type.