Local AutoML for CSV files

Why CSV remains the starting point for many ML workflows

Spreadsheets and CSV exports are everywhere. Sales teams export opportunities, operations teams export tickets, schools share teaching datasets, researchers collect tabular observations, and founders often begin with a spreadsheet before they have a data warehouse. CSV is not glamorous, but it is readable, portable, and easy to move between tools.

The challenge is that CSV modeling can go wrong quickly. A column may be an ID, a timestamp may define the split, a category may be too rare, a target may leak into another field, or missing values may appear mostly later in the file. Local AutoML should make these issues visible rather than turning the file into a black-box score. MLdeck treats CSV as a workflow, not just an upload step.

Local AutoML workflow in MLdeck

MLdeck starts by reading the CSV in the browser. The app profiles columns, estimates types, counts missing values, and lets the user review which fields should be included. The user then selects the target column and trains candidate models for tabular classification or regression. During normal browser training flows, raw CSV training data is not uploaded to a cloud training service.

This local workflow is useful when you need a first answer quickly: is there signal, which columns matter, which task type fits, and what warnings should be addressed before deeper validation? It is also useful in classrooms because learners can see the relationship between data shape, preprocessing, model choice, and evaluation evidence without installing Python.

Profiling, target selection, and preprocessing

Good CSV modeling depends on schema decisions. Numeric columns may need imputation, categorical columns may need encoding, outliers may need clipping, and text-like fields may need exclusion. MLdeck exposes these choices through a visual workflow so users can understand what is included in training. Target selection is treated carefully because the target defines whether the workflow is classification or regression and whether the evaluation evidence is meaningful.

Users should review ID columns, timestamps, post-outcome fields, and any column that would not be known at prediction time. Local AutoML can accelerate modeling, but it cannot automatically understand the business meaning of every field. Human review remains part of the workflow.

Baselines, leaderboard, and warnings

MLdeck compares models and presents leaderboard evidence, but a leaderboard is not the whole story. Baselines help users understand whether a model beats simple alternatives. Warnings help flag data quality issues, possible leakage, task ambiguity, and validation gaps. These signals are designed to make early evaluation more honest.

Results should be treated as exploratory unless strict validation is run. For important decisions, holdout strategy, temporal splits, class balance, fairness considerations, and external validation all matter. MLdeck is an MVP and early beta, so the responsible workflow is to use it for learning and iteration, then validate before relying on outputs.

Warnings are also useful teaching moments. They help explain why an easy-looking CSV can still produce fragile evidence when rows are duplicated, labels are delayed, or features describe events that happen after the prediction point.

Export options after CSV model training

After training, MLdeck can produce exportable ML artifacts for validation and deployment testing. 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 support review and documentation. Python artifacts can help technical users inspect or extend the workflow.

Exports should be handled carefully. If a model was trained from sensitive data, the model and report may still reveal useful information about that data. Share exported artifacts only with the same care you would use for other derived analytics assets.

Local CSV AutoML is also a good place to document assumptions. Note which columns were excluded, why the target was chosen, whether rows were sorted, what warnings appeared, and whether the data represents the future use case. Those notes often matter as much as the first model score because they explain whether the experiment should continue.

Local CSV AutoML FAQ