Review pipeline results and model reports

Review the result for the task family#

• Classification can show compact Test Results at the default 0.5 threshold, plus threshold-independent metrics and other operating points in the analysis views, ROC and precision-recall curves, a threshold sweep, confusion matrix, per-label metrics, label distributions, and SHAP explanations.
• Regression can show predicted-versus-observed fit, scatter and distribution views, residuals, and sample-level comparisons.
• Segmentation can show IoU and Dice summaries, per-class and per-sample distributions, pixel confusion, representative errors, and image or volume overlays.
• Detection can show precision-recall and AP50 summaries, threshold sweeps, background-aware confusion, IoU distributions, and representative detections.
• Sample evidence can include input, prediction, and ground truth for images, video frames, NIfTI slices, or TIFF data, plus Grad-CAM or Integrated Gradients when configured and supported.

Interpret metrics and operating points#

1. Confirm the selected version

   Use the version selector and verify the dataset, configuration, and completion state before reading the headline metric.

2. Read compact binary Test Results at 0.5

   For binary classification, compact Test Results reports its scalar evaluation metrics at the default 0.5 threshold. It does not represent a validation-selected or test-sample-optimal operating point.

3. Start with threshold-independent evidence

   For classification, ROC AUC and PR AUC summarize ranking across thresholds. They do not define the deployed decision threshold.

4. Identify analysis operating points

   Use the result analysis views to read sensitivity, specificity, precision, recall, F1, and confusion counts at each stated validation-selected or test-analysis operating point.

5. Separate validation selection from test analysis

   A threshold selected on validation data is a candidate operating point evaluated on test data. A test-sample-optimal threshold is diagnostic and should not be reported as a preselected deployment threshold.

6. Inspect class and sample behavior

   Use per-label, residual, IoU, confusion, error, and representative-sample views to find failure modes hidden by an average.

7. Read explainability as supporting evidence

   Use SHAP, Grad-CAM, or Integrated Gradients to investigate model behavior, not as proof that a prediction is correct.

Compare versions and create a model report#

1. Compare versions in the lineage

   Use current-versus-previous views to identify changes in configuration, headline metrics, detailed plots, and sample behavior.

2. Compare compatible pipelines

   Use Compare Pipelines for completed models with comparable datasets, targets, metric definitions, and operating points.

3. Open the model report

   When model-card reporting is available, generate or open the report for the selected model version.

4. Edit the narrative

   Verify intended use, dataset context, performance, limitations, and explainability notes against the displayed evidence. Remove sensitive data.

5. Preview and download

   Preview the rendered report and download it only after confirming that it names the intended version and does not overstate validation.