Params, Metrics and Artifacts

A simple rule

• Param = an input to training that does not change during a run.
• Metric = a measurement that may change during a run (per epoch / step).
• Artifact = a file the run produces (the model, plots, the confusion matrix, the SHAP report).

Common mistakes

1. Logging the dataset size as a metric. It's a param — it does not change during the run.
2. Logging per-epoch loss as a param. It's a metric with step=epoch.
3. Logging the model file as a metric. It's an artifact.

Step-aware metrics

for epoch in range(epochs):
    mlflow.log_metric('train_loss', loss, step=epoch)
    mlflow.log_metric('val_loss',   v_loss, step=epoch)

The MLflow UI then renders proper learning curves automatically.

Cooking analogy: params are the recipe ingredients (200g flour), metrics are the readings during cooking (oven temperature each minute), artifacts are the cake itself plus the photo and the receipt.