Train ML models in your browser

Why train ML models in the browser

Training in the browser changes the first step of machine learning. Instead of installing Python, creating a notebook environment, setting up packages, or uploading a CSV to a cloud service, a user can open a web page and begin. For many early workflows, the first goal is not a final model. The first goal is to understand the data: which columns are useful, which target makes sense, what warnings appear, and whether a simple model has signal.

Browser ML training is useful for people who need fast feedback. Analysts can test ideas. Students can learn by doing. Developers can create export artifacts for validation. Data scientists can compare quick baselines before writing custom code. During normal browser training flows, raw CSV data is not uploaded to a cloud training service. That makes the workflow approachable for quick, careful experiments.

Step-by-step browser ML workflow

A typical MLdeck session begins with a CSV upload in the browser. The app profiles the dataset, summarizes columns, and helps the user review feature choices. Next, the user selects a target column and confirms the task. MLdeck can then train candidate models for tabular classification or regression, depending on the target and schema.

After training, the user reviews leaderboard evidence, metrics, warnings, and export readiness. This sequence is meant to make modeling decisions visible. It encourages the user to ask whether the target is valid, whether important features are missing, whether the first rows are biased, whether class balance changed later in the file, and whether the evaluation strategy matches the real use case.

What algorithms and tasks MLdeck supports

MLdeck focuses on tabular classification and regression workflows. It can compare several scikit-learn style model families, including linear models and tree-based candidates where supported by the browser workflow. The available choices may depend on task type, selected features, data size, browser resources, and product maturity.

This focus is intentional. CSV AutoML is most useful when the problem is structured: predict a category, estimate a numeric outcome, compare feature influence, or evaluate whether a table has learnable signal. It is not a general deep learning workbench, image trainer, large language model trainer, or replacement for every cloud ML platform.

That narrower scope helps keep the workflow understandable. Users can see the target, features, preprocessing, metrics, and export contract in one place.

How evaluation evidence should be interpreted

Metrics are evidence, not permission slips. A high score may come from leakage, sorting, repeated entities, class imbalance, or a target that is easier in the sample than in the future. A low score may mean the data lacks signal, the target is noisy, the split is wrong, or preprocessing needs review. MLdeck surfaces warnings and evidence to support better judgment.

Because MLdeck is an MVP and early beta, strict validation should be used before relying on results for important decisions. Consider representative holdouts, temporal validation, external test data, parity checks for exports, and domain review. Browser training is excellent for learning and early evaluation; rigorous deployment decisions need more evidence.

Users should also ask whether the model will see the same kind of rows later. If the CSV was sorted by time, filtered to successful cases, or collected from a narrow period, the metric may overstate how the model behaves in a broader setting.

Exporting models and reports after training

After training, MLdeck can create exportable artifacts. ONNX exports are designed for portable ONNX Runtime inference, subject to parity validation. Docker packages can help test an API-style serving path. PDF reports can document metrics, warnings, and preprocessing. Python artifacts can help technical users inspect or extend the workflow.

Exports are useful because they turn an in-browser experiment into something that can be checked elsewhere. They should still be treated carefully. Validate schema, representative rows, error handling, and runtime behavior before using exported artifacts in important systems.

Browser ML training FAQ