Datasets

The Datasets page manages the datasets an experiment can run against.

​

Create a dataset

Datasets → New dataset with a name and description.

​

Add items

Open a dataset and add items:

• Input — JSON or plain text.
• Expected — the reference output (optional, used by dataset scoring).

Items are versioned with the dataset, so each experiment records exactly which version it ran against.

​

From traces

Capture real production runs from the Traces page into a dataset to build regression suites out of cases that actually happened.