Edit in dashboard

Add rows, edit Input / Expected / Metadata in the row panel, tag rows, and assign owners — all from the dataset detail page.

The dataset table is a read-only view of your test cases. Click any row to open the row panel on the right, then edit in the Fields tab. Changes apply to the live draft until you click Save.

Open a dataset

  1. Go to Datasets in the sidebar.
  2. Click a dataset name (or create one with + New dataset).
  3. The detail page shows the row table and toolbar.

The row panel is collapsed by default — it only opens when you click a row. Close it with the X in the panel header to return to full-width table view.

Add rows

  • Click + Row in the toolbar to append an empty row at the bottom.
  • If the dataset is empty, click Add first row below the table.

Edit fields

With a row selected, use the Fields tab in the row panel:

Input, Expected, Metadata

Each field has a view type dropdown and a toolbar:

ViewEditable?Use it for
YAMLYesHuman-friendly editing of structured data.
JSONYesPrecise editing; validates on blur.
TextYesPlain string values (especially Expected).
HTMLRead-onlyPreview HTML strings or escaped JSON.
PrettyRead-onlyFormatted JSON preview.
TreeRead-onlyIndented key/value tree view.

Toolbar actions:

  • Copy — copy the current view to clipboard.
  • Word wrap — toggle line wrapping.
  • Reformat — re-indent YAML or JSON.
  • Download — save the field content as a file.
  • Split view — editor on the left, Pretty preview on the right.

Edits save when you blur the field (click outside the editor). Validation errors appear below Input and Metadata if YAML/JSON is malformed.

Input must be an object — a JSON/YAML map of { variable: value }. In YAML or JSON view, arrays and bare scalars are rejected. In Text view, a bare string is auto-wrapped as { value: "your text" }.

Metadata must be an object (or empty) in YAML/JSON; empty becomes {}.

Expected can be any string, including JSON-serialized objects.

Tags

  • Click the Tag icon in the panel action bar (or Tag in the top header when a row is selected).
  • Add, remove, or pick from tags already used in the dataset.
  • Tags appear as chips in the table's Tags column and can be filtered via the All dataset rows view dropdown.

Assign

  • Assign — pick a team member as the row owner (shown in the Assigned column when Display is on).

Toggle Display in the toolbar to show or hide collaboration columns (Assigned, Origin).

Table toolbar

ActionWhat it does
ImportUpload a file
+ RowAdd an empty row
All dataset rows viewFilter by tag
ResetDiscard unsaved changes and reload from server
Save asCreate a snapshot
Add filterText search across row values
DisplayToggle collaboration columns
DownloadExport CSV

Field schemas (derived columns)

Open the derived-column UI from Field schemas in the top bar, or hover the # column header and click +.

Each derived column is computed from existing row data and shown read-only in the grid; values recalculate when source data changes. Choose a mode when creating one:

ModeWhat it does
BasicPick a field path to display (e.g. input.query, metadata.category, expected_output).
ExpressionCompute from fields with a formula — LENGTH(), UPPER(), LOWER(), `

Examples: LENGTH(input.query), UPPER(metadata.category), input.query || " -> " || expected_output.

Row panel tabs

TabPurpose
FieldsEdit Input, Expected, Metadata, tags, assignee
RunsExperiment results that ran on this row
ViewsSaved filter views (coming soon)

Save and snapshots

  • Save (top bar, appears when there are unsaved changes) — persists the draft to the server.
  • Snapshots — freeze the current rows as an immutable numbered version. Pin a snapshot when running an experiment for reproducible comparisons.

Run an experiment

Click Evaluate → Run in Experiment to batch every row through a prompt and model(s). See Experiments.

Next

On this page