Grouped Metadata Plot

The Grouped Plot lets you compare distributions across experiment groups.

Instead of comparing single values, you compare full distributions using box plots and summary statistics.

Each row represents a group of selected executions.

Selecting Executions

On the left side panel, you can:

• Add executions to the comparison

• Remove executions

• Search and filter runs

Only the selected executions appear in the grouped comparison.

Define Your Grouping

In the right panel, configure:

Group By (rows)

Choose a categorical metadata key. Each unique value becomes a row.

Examples:

• dataset_name

• model_version

• instance_type

• architecture

Values (columns)

Select one or more numeric metadata fields to analyze.

Examples:

• val_accuracy

• val_loss

• training_time

• latency

Each selected value becomes a column with its own box plot and statistics.

Visible Aggregates

You can toggle which statistics to display:

• n (number of executions)

• Min

• Q1

• Median

• Q3

• Max

• Mean

You can also:

• Show / hide box plot

• Show / hide individual observations

Example Configurations

Compare Datasets

• Group by: dataset_name

• Values: val_accuracy, val_loss

See which dataset performs best and which introduces more variance.

Compare Model Versions

• Group by: model_version

• Value: val_accuracy

Understand which version is most stable and which has higher median performance.

Infrastructure Benchmark

• Group by: instance_type

• Value: training_time

Detect variability or long-tail performance differences.

Why Use Grouped Plot?

Use it when you want to:

• Compare distributions (not just averages)

• Detect instability within experiment groups

• Identify variance patterns

• Evaluate A/B tests or multi-variant experiments

• Benchmark datasets, architectures, or infrastructure

It's especially useful when you care about consistency, not just peak performance.

How To Use It

1. Select executions in the left panel.

2. Open Metadata view.

3. Click Grouped Plot.

4. Choose:

   • Group By key

   • One or more Values

   • Visible aggregates (optional)

5. Apply filters if needed.

The table and box plots update instantly.