Train: from reviewed data to model
Export a reviewed project, validate the output in your training loader, and record the data and code behind each run.
A training-data handoff needs more than a download. Confirm which examples are included, whether they passed your quality gate, and how the annotations map to the targets your model expects.
Avala provides the data and export workflow. You run training and evaluation in your own stack and connect the results to the data release you used.
Create an export
Set AVALA_PROJECT_UID to the existing project you intend to export. The following command creates an export job:
avala exports create \
--project "${AVALA_PROJECT_UID:?Set AVALA_PROJECT_UID to your project UID}"
Copy the export UID returned by the command into AVALA_EXPORT_UID, then wait for the job:
avala exports wait \
"${AVALA_EXPORT_UID:?Set AVALA_EXPORT_UID to the returned export UID}"
Check the returned status before using the download. A terminal status can be a failure as well as a completed export. These CLI commands accept a project or dataset source; they do not accept a format-selection flag.
Validate the training input
Inspect the exported schema and use the conversion or loader required by your framework. Validate a small sample before a full training run:
- Do class identifiers and label meanings match the project specification?
- Are coordinates, units, and timestamps interpreted correctly?
- Are the intended quality states and selected examples included?
- Are training and evaluation partitions separate?
An export does not certify that the model is safe or that every label is correct. Those conclusions depend on the review process and your evaluation.
Record what changed
Keep the export UID, project configuration, preprocessing code, model configuration, and evaluation results together. If a metric regresses, this record helps distinguish a data change from a code or training change.
Webhooks and job orchestration can connect an export to downstream processing where supported and configured. Make the quality and evaluation gates explicit before automating the handoff.
Bring the result back to the data
Look beyond an aggregate score. Identify the scenarios that improved and the failures that remain. That comparison tells you whether the next round needs more examples, a revised label specification, or a different model approach.