Training Finished. Now What?
A completed training run produces a model, but that model is only useful if it performs well on data it has not seen before. Evaluation is where you find out whether the model actually works or whether the dataset, architecture, or hyperparameters need adjustment.
What This Tutorial Covers
Datature walks through the model evaluation tools built into Nexus:
- Reviewing training metrics: loss curves, mAP, precision, and recall
- Running inference on validation images to spot failure cases
- Using the confusion matrix to identify which classes the model confuses
- Comparing metrics across multiple training runs
The tutorial covers these evaluation workflows in about three minutes.
What to Look For
High overall accuracy can hide problems. A model might score 95% mAP on common classes but miss rare defect types entirely. The confusion matrix shows these patterns clearly. Loss curves that plateau early or spike late signal training issues that require different learning rates or more data. Datature surfaces all of these metrics in one dashboard so you do not need to parse log files or write evaluation scripts manually.
For a deeper explanation of confusion matrices, read How to Evaluate Computer Vision Models with Confusion Matrix.
