Evaluation Dataset

Manage evaluation examples with inputs, expected outputs, labels, source, difficulty, coverage, and review state before running model tests. This example brings Examples and Labels together so AI product teams can find the next task without losing context.

It includes representative records such as Permissions edge-case prompt and Billing refund scenario. Use the screenshot to study the layout, then adapt its components and product states to your own design system.

Evaluation Dataset

Template facts

Screen

Table

Primary action

Add example

Pattern

Table

State

Default, loading, empty, error, permission, success

Design system

Configurable product UI theme

What this screen helps users do

Product teams need to add example while keeping examples, labels and exceptions visible. The interface should show what changed and who should act next.

Key elements

  • Main workspace: Examples gives users an obvious place to start.
  • Supporting view: Labels keeps the related context close.
  • Primary action: Add example stays visible when people need it.
  • First signal: Examples shows a useful value (842).
  • Second signal: Labels gives additional context (18).
  • Sample record: Permissions edge-case prompt.
  • Next record: Billing refund scenario.

States worth designing

  • No rows in current filters: say what changed, protect progress, and offer a clear next step.
  • Sorting or bulk action in progress: say what changed, protect progress, and offer a clear next step.
  • Row update conflict: say what changed, protect progress, and offer a clear next step.
  • Bulk update partly failed: say what changed, protect progress, and offer a clear next step.
  • Restricted columns: say what changed, protect progress, and offer a clear next step.
  • Export complete: say what changed, protect progress, and offer a clear next step.

Where this screen sits in the flow

Examples → Labels → Add example → Review outcome

Patterns used

Table, Examples, Labels, contextual actions, error recovery

Why these choices work

Prioritize columns that help people act. Keep filters when readers drill into details and return; show who owns each unresolved exception.

Pay attention to examples and labels. People need to distinguish Permissions edge-case prompt from Billing refund scenario. Keep examples visible, especially when a record changes owner or status.

Before this screen reaches handoff, verify keyboard focus, empty and permission states, and what happens after someone chooses “Add example”. Make error recovery as clear as the happy path.

Related screen examples to explore next: Prompt Versioning, Human Review Queue, Retrieval Playground.

Start from this in Figr

Design a evaluation dataset screen inside our existing product. Use our current components, tokens and content language. Include Examples and Labels views, credible records like Permissions edge-case prompt, and a clear “Add example” action. Use our product permissions and show loading, empty, error and success states. Keep the output editable in Figma.

Open Figr