Läufe in der MLflow-Oberfläche vergleichen

Open an experiment in the UI and you get a runs table: one row per run, with its params and metrics laid out as columns side by side.

Click a metric column header to sort. Order by accuracy descending and your best run jumps straight to the top of the table.

Type a query in the search box to filter. Keep only the runs that match a condition, like a metric above a threshold you care about.

Too many columns? Toggle which params and metrics are visible so the table shows only the few that actually drive your decision.

Tick the checkboxes next to a few runs and hit Compare. MLflow opens a focused view of just those runs side by side.

The parallel coordinates plot draws each run as a line across params and metrics, so you can spot which settings lead to high scores.

A scatter plot maps one param against a metric. It reveals trends, like accuracy rising then falling as tree depth grows.

Click any run to open its detail page. There you see every logged param, metric curve, tag, and artifact for that single execution.

On the detail page, stepped metrics render as curves. A loss that flattens early hints you trained longer than you needed to.

You are not stuck in the browser. Use search_runs to pull matching runs into a DataFrame and analyze them programmatically.

Comparing is how you choose a winner. Sort, filter, and plot until one run clearly stands out, then move it toward production.

Let us check how you surface your best run in the UI.

You sorted, filtered, and plotted runs, then opened details to confirm a winner. The UI turns a pile of runs into a clear choice. ✅