Annotate, review & publish

Define consistent labels

Describe what counts as each class, how to handle occlusion, and which objects to ignore. Check a small set together before labeling the full dataset. For VLM examples, define the expected answer style and avoid contradictory answers for the same question.

Label and correct

1. Open the project’s annotation dataset or workbench.
2. Inspect the image and its labels. Use the manual tools available for the dataset type.
3. If AI-assisted annotation is configured, run it on a small batch first and inspect its output.
4. Correct missing objects, incorrect classes and poor boundaries; save your changes before leaving the image.

AI output is a starting point. Human inspection is needed for missed objects, rare cases and incorrect descriptions.

Review the dataset

Use the review controls shown by your installation. Check class consistency and a selection of difficult images. Read any approval confirmation before submitting it. Review requirements differ by workflow; do not assume an AI task’s completion automatically approves the data.

Publish a training version

Publish through the project’s dataset workflow when it is ready. Record the version identifier and check it appears among the training inputs. If a published version needs corrections, prepare a new version rather than assuming an existing run will use later label edits.

Generated images

When image generation is configured, inspect generated samples for visual artifacts and whether the target objects are usable. Keep their origin identifiable. Generated images do not replace an independent real-world evaluation set.

Next: configure training.