Evaluate & export

Evaluate against the task

Use samples that were not used for training. Include common scenes and difficult cases such as small objects, occlusion and changing illumination. Keep the same dataset and settings when comparing models.

Run an evaluation when configured

1. Open the evaluation area for a model or project.
2. Select the model and an evaluation dataset you are allowed to use.
3. Check the task type and available metrics, then submit the evaluation if the service is ready.
4. Wait for its final status. Inspect failed examples as well as aggregate scores.

Metric availability depends on the evaluation workflow. A score from another dataset or model configuration is not a comparable improvement. For VLM tasks, inspect answer quality alongside the configured evaluation criteria.

Keep the lineage

Record the model identifier, training input version, evaluation input and settings. If results are poor, classify the errors and correct the data or model setup before another run.

Download and conversion

Use the export/download controls available for the selected artifact. Wait for the export task to finish, then verify the downloaded package can be opened and contains the expected files. Refresh the export history if a completed result has not appeared.

Conversion depends on architecture and target runtime. An uploaded checkpoint is not automatically convertible, and conversion success is not a deployment acceptance. Confirm the required format, runtime and hardware before preparing a large package.

Deployment handoff

Agree on a separate deployment and acceptance plan for the target device or service. Preserve the previous working model until the replacement has passed its own checks.