Interpreting Training Results
After training, Neuro QP shows both the classifier output in the viewer and short hints about the training run.
The main goal is to use that information to decide what to do next.
Start with the classifier output
The most important quality control is still checking a few slices visually.
Look at whether the ON/OFF calls make sense in practice.
This matters more than any single summary metric.
Use the hints as guidance, not as the goal
The hints are there to point you toward likely weaknesses in the current model.
They help you decide what samples to add next.
They should not be treated as the final definition of model quality.
Accuracy is only one hint
Classifier accuracy or F1 is useful, but it is not the main thing to optimize by itself.
The real goal is a classifier that behaves well on the real slices used in the analysis.
Why scores can go down even when the model improves
As you add more samples, the model usually gets better.
But the test set also becomes larger and harder.
This is especially true when you use error-mining and deliberately add difficult examples.
Because of that, scores can go down even while the model becomes more useful in practice.
Use the result to decide the next step
If the output looks wrong in a specific way, add samples that target that failure mode.
Use the hints and the slice review together to decide what to correct next.
When the model is good enough
A model is good enough when the results look trustworthy on a few representative slices.
The goal is not to maximize a number, but to reach results that are reliable enough for the analysis.