What model accuracy doesn’t tell you
Go beyond a single number. Explore evaluation metrics, test sets, and practical failure analysis.
One number hides the failures
A single accuracy score averages over every kind of example. Class imbalance can make a useless model look excellent, so report precision, recall and per-class results alongside it.
Respect the test set
Keep a held-out set that you never tune against, and make sure it reflects production data rather than the training distribution. Leakage between splits is the most common source of inflated results.
Study the errors
Read a sample of wrong predictions by hand. Grouping failures into a handful of categories tells you more about what to fix next than any additional metric.
