F1 score combines precision and recall into one number using their harmonic mean. It's calculated as 2 times precision times recall, divided by precision plus recall.

‍

‍Why it matters

A plain average can hide a badly unbalanced system. With 100% precision and 10% recall, the simple average is 55%, which sounds decent. F1 comes out around 18%, much closer to the weaker number, which is a fairer picture of a system that misses most of the real cases.

‍

‍When to use it

‍F1 is handy when you need one number to compare models quickly and both kinds of mistakes matter about equally. When one is clearly worse, like missing a safety violation, look at precision and recall separately or use F-beta, because F1 alone doesn't show which way the system is failing.

‍

‍Variants worth knowing

F1 weights precision and recall equally. The more general F-beta score lets you tilt it: F2 puts more weight on recall, and F0.5 puts more weight on precision. For datasets with several categories, macro F1 averages the score across categories equally, while micro F1 pools all the items together, so large categories dominate it.