How do I add examples to a task?
Upload inputs and expected answers in the task’s Standard tab as CSV, JSON, or JSON Lines. Each upload creates a new dataset version. Comparisons keep the version they tested, so adding examples does not rewrite earlier results or change a live release.
Prepare representative examples
Examples are inputs with the answer your reviewers expect. Upload a CSV, JSON, or JSON Lines file; in a CSV the first column is the input and the other columns are the answer fields. Each upload makes a new dataset version, and the previous one stays. A comparison names the dataset version it ran on.
Known organization credentials and recognizable secret values are removed from examples. Ordinary fields such as token or session keep their names and value types. When content changes, the console asks you to inspect it before comparing models.
Examples also come from production: a reviewer’s correction, once promoted, becomes a new example, and an agent step can be added to the task’s examples. See Checking answers.
Include normal requests, difficult cases, and cases the task should refuse. Match each expected answer to the task’s answer shape. Inspect any protected-value warning before starting the comparison.