Training Diagnostics Checklist | Aabha AI Academy

Use this when a model trains badly, too well, or unpredictably.

First Checks

[ ] Did the notebook run from top to bottom?
[ ] Are random seeds visible?
[ ] Are train, validation, and test roles separate?
[ ] Do input tensors have the expected shape?
[ ] Does the target have the expected shape and type?
[ ] Is the loss function suitable for the target?
[ ] Is the optimizer updating model parameters?
[ ] Is optimizer.zero_grad() used before backpropagation?

Curve Checks

[ ] Training loss decreases.
[ ] Validation loss is measured separately.
[ ] Training improves but validation does not: possible overfitting.
[ ] Neither improves: possible underfitting, wrong learning rate, data issue, or code bug.
[ ] Loss jumps around: possible learning-rate or data-scale problem.

Data Checks

[ ] Labels look plausible.
[ ] No duplicate leakage between train and validation.
[ ] Features are scaled or encoded where needed.
[ ] Class balance or target range is inspected.
[ ] Baseline result is known.

Decision

[ ] Accept the result only as a learning artifact.
[ ] Revise the next experiment clearly.
[ ] Record the defect and retest.
[ ] Write a limitation before sharing the notebook.