Skip to main content

Finalize Dataset

Finalization validates a draft, generates its dataset artifacts, and locks that version for reproducible model training.

Finalization checklist​

Before selecting Finalize, confirm that the draft has:

  • at least one class
  • target wavelengths
  • a data processing level
  • Training images or Training reference spectra
  • a data assignment method and valid bucket contents
  • a sampling method and class split type
  • a number of latent variables if PCA is enabled
  • no unresolved dataset-validation errors or incompatible assigned entries

Only finalized labels for the selected classes contribute image samples. If label attributes are selected, the included labels must supply all of them.

Check bucket design​

On Images, confirm the assignment method and split:

  • Manual uses separate Training, Validation, and Test groups.
  • Automatic splits the Training & Validation group according to the slider; the class split type determines whether the split unit is a pixel or label.

Keep evaluation data separate from training whenever you use its metrics to estimate generalization. Automatic Pixel splitting can put correlated pixels from the same label and image in both Training and Validation. Label splitting keeps a label region together, but the buckets can still share an image. For an independent scene-level measure, use Manual assignment and hold out whole images, sites, dates, sensors, or acquisition conditions that represent the deployment domain.

The correct sample counts depend on the model and deployment domain; Clarity does not prescribe a universal pixel count. A class with only one eligible sample cannot support independent Training and Validation. The current Automatic split can duplicate that sample across both buckets, so add independent data instead of interpreting the resulting metric as held out.

Review analysis when available​

The Analysis area can show class/bucket counts, spectral statistics, and PCA plots when those artifacts exist. Use them to investigate imbalance, spectral overlap, distribution shifts, and outliers. They are diagnostic evidence, not an automatic guarantee that a dataset is suitable.

With Alpha features enabled, Analyze can generate analysis for a valid draft before finalization.

Finalize​

  1. Return to Overview and resolve any validation alert.
  2. Select Finalize.
  3. Follow the finalization job status.
  4. After it completes, set the finalized version's publication status to Unpublished, Published, or Deprecated as appropriate.

To change a locked result, create a new version or clone it into a new dataset.