How to use MLOps Desktop

This page is the everyday user manual. It explains what to click, what to check, and how to decide the next step after each run.

MLOps Desktop is most useful when you already have a CSV and want a repeatable local workflow for training scikit-learn models. It does not replace every notebook or production platform. It helps you move the common loop into one place: load data, train, evaluate, compare, explain, export, and serve.

The normal workflow

Most pipelines follow this shape:

DataLoader → DataSplit → Trainer → Evaluator

Add these only when you need them:

Trainer → Model Exporter
Trainer/Evaluator → Explainability, Runs, Models, Serving panels

Use this rule of thumb:

If you need to	Start with
Train a model from a CSV	DataLoader → DataSplit → Trainer → Evaluator
Try a better configuration	Change the Trainer, run again, compare in Runs
Search many configurations	Set Trainer to Tune, then watch Trials
Understand model behavior	Open the Metrics run and click Explain
Save a model file	Add Model Exporter after Trainer
Test an HTTP prediction endpoint	Use the Serving and playground panels

1. Start with a clean CSV

Before building nodes, check the data file once:

• The target column exists and has the exact name you plan to select.
• The first row contains headers.
• Empty values are expected and not a file export mistake.
• Very high-cardinality text columns, IDs, and free-form notes are either removed or intentionally kept.

Tip

The Trainer handles common preprocessing such as missing values and categorical encoding, but it cannot fix a target column that is missing, mislabeled, or not meaningful.

2. Load and inspect data

1. Add a DataLoader node to the canvas.
2. Select the node, then choose your CSV file.
3. Run the pipeline once with only the data node if you want to confirm the file loads.
4. Open the Data Profile tab and check row count, column names, types, and missing values.

If the profile looks wrong, fix the CSV before adding model nodes. Bad column names or unexpected empty columns are faster to fix at this stage.

3. Split before training

Add DataSplit between DataLoader and Trainer for a clearer train/test boundary.

Recommended starting values:

Setting	Start with	Notes
Test split	0.2	Use 20% of rows for evaluation
Random state	42	Keeps runs comparable
Stratify column	target column	Use for classification when classes are imbalanced

For very small datasets, a large test split can leave too little training data. If you see unstable metrics, try a smaller split or use cross-validation during tuning.

4. Train the first baseline

Add Trainer, connect it after DataSplit, and configure:

Setting	Classification baseline	Regression baseline
Model type	Random Forest Classifier	Random Forest Regressor
Target column	Your label column	Your numeric target column
Mode	Train	Train

Run the pipeline and watch the Logs tab. A good first run is not about getting the best score. It is about proving that the data, target, and model configuration work end to end.

5. Evaluate before changing things

Add Evaluator after Trainer, run again, then review the Metrics tab.

For classification, check:

• Accuracy for a quick overview.
• Precision, Recall, and F1 when false positives or false negatives matter.
• The confusion matrix to see which classes are being mixed up.

For regression, check:

• R² for overall fit.
• RMSE when large errors should hurt more.
• MAE when you want an error in the same unit as the target.

Caution

Do not tune hyperparameters until the baseline makes sense. If the target is wrong or leakage exists, tuning only makes a bad setup look more polished.

6. Compare runs instead of trusting one score

Each run is saved in the Runs panel. Use it to answer practical questions:

Question	What to compare
Did the new model improve?	Same dataset, same split, different model type
Did tuning help?	Baseline run vs tuned run
Is the score stable?	Same config across repeated runs or cross-validation
Is the model worth keeping?	Metrics, notes, tags, and explainability outputs

Use notes and tags on runs when you change something important, such as baseline, rf-depth-10, tuned, or removed-id-columns.

7. Tune only after the workflow is stable

Switch the Trainer to Tune mode when you have a working baseline and want to search better parameters.

Start small:

Setting	Practical starting point
Search strategy	Bayesian / TPE
Trials	20 to 50
CV folds	3
Metric	accuracy or f1 for classification, r2 for regression

Watch the Trials tab while it runs. If trials are slow, reduce the trial count, reduce CV folds, or choose a faster model before trying a bigger search.

8. Explain the model when you need confidence

After an evaluation run, use Explain from the metrics area to generate explanation outputs.

Look for:

• Feature importance: which columns drive the model most.
• SHAP outputs: how features push predictions up or down.
• Partial dependence plots: how the prediction changes as one feature changes.

Use explanations to catch obvious problems. If customer_id, row_number, or another accidental ID dominates feature importance, go back to the data and remove it.

9. Export or serve only the model you trust

When a model is worth keeping:

• Use the Models panel to register and compare versions.
• Promote stages only when the model has a clear purpose: Staging, Production, or Archived.
• Add Model Exporter if you need a .joblib, pickle, ONNX, or CoreML file.
• Use Serving and the playground to test local HTTP predictions before wiring the model into another app.

Keep the exported model and its metadata together. If feature order or package versions change, loading the model later can fail or produce wrong predictions.

What to do when a run fails

Check issues in this order:

1. Logs tab — read the first real error, not only the last line.
2. Python path — confirm the app is using the Python installation where packages are installed.
3. Packages — install missing packages such as scikit-learn, pandas, optuna, or shap.
4. CSV columns — confirm target and feature column names match exactly.
5. Model choice — confirm classification models are used for class labels and regression models for numeric targets.
6. Data size — confirm you have enough rows after splitting.

If you are stuck, copy the error from Logs and compare it with the troubleshooting guide.

A simple checklist before sharing a result

• The pipeline loads from a saved file.
• The same run can be repeated without changing the CSV path.
• Metrics are from a held-out split or cross-validation.
• The target column is correct.
• Obvious leakage columns were removed.
• Important runs have notes or tags.
• The model version has metadata and an export path if it will be used outside the app.

Next steps

Build the first pipeline Use a sample dataset and run DataLoader → Trainer → Evaluator.

Tune hyperparameters Use Optuna after you have a reliable baseline.

Deploy a model Start local serving and test predictions in the playground.