Metrics and acceptance criteria
Read a score with its task, reference, and decision rule.
The available metrics depend on the selected task and test. Choose them for the question you need to answer, and retain their settings with the result.
Common task-specific measurements
| Task | Examples in the metric catalog |
|---|---|
| Image or tabular classification | Accuracy, Precision, Recall, F1 score |
| Object detection | AP@0.50, AP@0.75, mAP 50–95, mAR, overlap measurements |
| Instance segmentation | Mask AP, Mask mAP 50–95, Mean Dice score, Mean IoU, Pixel precision/recall |
| Oriented boxes | OBB mAP 50–95 and size-specific OBB AP |
| Tabular regression | Mean absolute error, Root mean squared error, Mean absolute percentage error, R² |
The task’s picker is authoritative for the options available in a particular test. Compare scores only when the task, reference labels, settings, sample coverage, and conditions support that comparison.
Keep the criterion explicit
A criterion defines the requirement a result must meet. Record the reason for the chosen rule and its intended use in the Test plan. Error metrics and accuracy-style scores have different meanings; inspect the metric explanation before interpreting better or worse.
Preserve uncertainty and gaps
Review examples and coverage alongside aggregate scores. Incomplete tests and omitted requirements stay visible in the report and evidence review.