Dataset validation

Every dataset added to Luna Studio — uploaded, fetched, generated, or imported — goes through validation. Validation determines whether the dataset can be used in a run.

​

States

Validation has four possible end states: The state appears in the run creation flow when Luna Studio checks a selected dataset:

• The Selected dataset card in Step 2 and Step 3 of the run creation flow.

​

What Luna checks

​

Schema checks

• The file parses as CSV or JSONL.
• Required columns are present:
  • Test sets — metric-specific feature columns and label.
  • Training sets — metric-specific feature columns and label when the dataset is already labelled.
• Column types match the metric’s output type (e.g. labels are parseable as Boolean for a Boolean metric).

​

Content checks

• File encoding (UTF-8 expected).
• Row count > 0.
• Empty rows are flagged as warnings.
• Inputs that exceed the model’s max token limit are flagged.

​

File checks (uploads and URLs)

• File size within the upload limit.
• For URLs: the URL is reachable; the response content type is appropriate.

​

Common errors

​

Common warnings

​

Unlabelled training logs

If uploaded or imported training logs are missing labels, Luna Studio does not train on them directly. It opens the label-only generation flow, uses the selected metric prompt to create labels, and saves a labelled training dataset. Generated training sets are always labelled.

​

When validation fails mid-run

If a dataset that was previously Validated later fails (e.g. the originating Galileo dataset changed), the run that consumed it can fail with a validation error. See Run failed.

​

Re-validating

Luna validates a dataset once at add-time and once per run launch. There’s no manual “re-validate” button — to re-check a dataset, re-add it (or fetch the URL again).

​

Where to go next