Check Dataset quality
Use findings to decide what needs inspection or correction.
A quality scan describes the selected Dataset version. Open the Dataset’s Quality panel to follow the scan and inspect results and findings.
Read checks in context
Vision checks cover file integrity and format, exposure, contrast or resolution where applicable, aspect ratio, blur, target completeness, class balance, annotation validity and consistency, similarity, and split coverage.
Tabular checks include target completeness and import-time task consistency. Check applicability in the results rather than expecting every vision check to apply to tables.
Inspect the source samples
Use findings to find the affected data and examine representative examples. Decide whether the issue changes the evaluation question, needs corrected labels, or should be documented as a limitation.
A scan provides evidence about data quality; corrections are a separate action. Use Edit labels for annotation changes and Update data for a new data version.
Keep the test basis explicit
When preparing a Run, select the intended Dataset version. If quality limitations remain, retain them in the plan and report.