Datasets

After Quickstart, open Data → Datasets to choose the task collection for your next Job. A Dataset is versioned input, not a running Job and not a training checkpoint.

Choose reproducible input

1. Select your workspace and project, then search Datasets by name.
2. Open the Dataset and inspect its available revisions, configs, splits and Tasks.
3. Record the selected name and version. Submit a Dataset Job with that selection; choose a task_id when running just one problem through a supported Runner submission.

The catalog currently leaves unavailable visibility, owner and other metadata as —. This does not mean public visibility or zero Tasks. An empty search result only describes the catalog visible to your connection.

When a submission omits a version, config or split, the platform resolves and freezes its default selection. Record the resolved input from the response for reproduction; changing a catalog default does not alter an existing Job. Dataset availability and execution support must be checked on the target platform.

Keep inputs separate

Dataset Tasks contain the problem data and evaluation definition. A Runner defines the entrypoint, and a Profile supplies reusable parameters. Credentials belong in API Key configuration, never Dataset files or Profiles. For training samples, follow the dataset setup in SFT quickstart.