Leaderboard

The leaderboard compares the evaluation performance of multiple models under the same evaluation dimensions, calculating 0-100 scores based on the overall scores of the associated evaluation dimensions and ranking them.

Leaderboard overview

The leaderboard is used to compare the evaluation performance of multiple models under the same evaluation dimensions, helping you intuitively judge the strengths and weaknesses of each model.

Create a leaderboard

On the model evaluation page of the Model Studio console, select the Leaderboards tab and click Create Leaderboard. Enter a leaderboard name (up to 50 characters), select evaluation dimensions (required, cannot be changed after creation—delete and recreate if the wrong one is selected), and optionally associate existing evaluation tasks (completed tasks using the same evaluation dimensions). Tasks already bound to the leaderboard cannot be added again.

Leaderboard scores and usage

Leaderboard scores range from 0 to 100 and are calculated based on the overall scores of the associated evaluation dimensions, which differs from the scoring range of the evaluation dimensions themselves (such as 0-5). The top three on the leaderboard are distinguished by color markers: gold for 1st place, silver for 2nd place, and bronze for 3rd place.

You can create evaluation tasks directly on the leaderboard details page (see Create a custom evaluation task for the creation workflow). When entering from the leaderboard entry, the evaluation dimension and leaderboard fields are automatically locked to ensure that the new task uses the same scoring standards as the leaderboard.

To ensure fair model comparison, it is recommended that all evaluation tasks in a leaderboard use the same evaluation datasets.

Leaderboard management

Deleting a leaderboard does not affect the associated evaluation tasks and evaluation results. Removing a task from a leaderboard is reversible—after removal, the task can be added again. Tasks associated with a leaderboard must be removed from the leaderboard before they can be deleted.