Skip to main content
Datasets are tables of examples. They can be written by hand, imported from CSV, created from production traces, or generated for broader coverage. Use the Datasets section for the current dataset workflow.

Dataset boundaries

Datasets are project-scoped. A dataset can be linked to prompts through evaluator configuration or dynamic prompt columns. Keep regression and golden datasets close to the prompts they protect so Improve and evaluation workflows can find the right evidence.

Library view

The project Datasets library shows dataset name, description, row count, column count, linked prompts, and update history. Use it to find:
  • Empty datasets that need rows.
  • Large datasets that may need splitting.
  • Datasets linked to active prompt evaluators.
  • Dynamic prompt columns that create prompt dependencies.
  • Recently changed datasets that could affect evaluation results.

Common columns

  • Prompt variables.
  • Expected outputs.
  • Labels or annotations.
  • Evaluator-specific inputs.
  • Metadata for filtering and review.