Dynamic Leaderboards in Evaluations

Dynamic Leaderboards let you configure, customize, save, and update Leaderboard views directly from an evaluation.

When working with Weave Evaluations, you can easily visualize and customize your experiment results as Leaderboards. Dynamic Leaderboard views automatically stay up to date as new evaluation runs are added.

Visualize Evaluation results in a Leaderboard

When your project contains Weave Evaluation data, you can use the evaluation table to quickly create a Weave Leaderboard view based on a filtered subset of results.

1. Navigate to wandb.ai.
2. In the Weave sidebar menu, click Evaluations.
3. Apply filters to the evaluation table to narrow the data to the models, datasets, or runs you want to compare.
4. In the evaluation table toolbar, click Visualize. Weave automatically creates a Leaderboard panel using only the data currently filtered in the table.
5. In the Leaderboard panel header, click Configure to open the Edit Leaderboard panel.
   The Edit Leaderboard panel gives you fine-grained control over how models, datasets, scorers, and metrics appear.

The following shows how a filtered evaluation table is visualized as a Leaderboard and where to configure the resulting Leaderboard.

Configure Leaderboard elements with visibility and custom names

The following shows the Edit Leaderboard panel with four configuration tabs: Models, Datasets, Scorers, and Metrics.

In the Edit Leaderboard panel, you can:

• Activate/deactivate display
  Select which models, datasets, scorers, and metrics appear in the Leaderboard by checking or unchecking them.

• Rename models, datasets, and scorers
  Assign display-friendly names (for example, renaming a model run to GPT-4 or a dataset to JokesV1).

  Renamed items:

  • Update immediately in the Leaderboard
  • Remain clickable so you can still open the underlying reference in the side panel
  • Automatically propagate anywhere the Leaderboard view is used

This makes it easier to compare experiments using meaningful, human-readable names without changing the underlying objects.

Configure Leaderboard metric behavior and coloring

In the Edit Leaderboard panel, for each metric, you can specify whether:

• Higher values are better, or
• Lower values are better

This setting directly affects Leaderboard coloring:

• Green highlights the better value.
• Red highlights the worse value.
• Colors automatically invert when you switch between “higher is better” and “lower is better”.

This ensures that visual cues remain accurate across different types of metrics (for example, accuracy vs. latency or error rate).

Save and reuse Leaderboard views

In the Edit Leaderboard panel, you can save your customized Leaderboard as a reusable view by clicking Save.

The saved Leaderboard view captures:

• Selected models, datasets, scorers, and metrics
• Renamed display labels
• Metric direction settings (higher or lower is better)
• Applied filters

Switch between saved views

Click the menu icon (☰) next to the Evaluations page title to open saved views. You can:

• Return to the default view to see the full dataset.
• Reopen a saved view to restore all customizations instantly.

When you reopen a saved view, all renames and metric settings are preserved.

---

Dynamic updates as evaluations change

Saved Weave Leaderboard views are dynamic:

• As new evaluation runs are added
• And as results match the saved filters

The Leaderboard automatically updates to include them, without requiring manual reconfiguration.

This lets you use views as living leaderboards that evolve alongside your experiments.