We use **Accuracy** as the primary performance metric. For each test image, if your model’s predicted class matches the true coin class, it counts as correct.