F1 score — Glossary

The longer version

The F1 score is the harmonic mean of precision and recall. Precision asks what share of predicted positive cases were truly positive; recall asks what share of the actual positive cases the system found. F1 rises only when both measures are reasonably strong.

For example, a detector that flags nearly everything may find most real problems but create many false alarms, giving high recall and low precision. A very cautious detector may have strong precision yet miss real problems, giving low recall. F1 summarizes that trade-off on a scale from 0 to 1.

It does not include true negatives directly, and it does not tell you whether a false positive is as costly as a false negative. In multi-class evaluations, micro, macro and weighted averaging can produce different results, especially when categories are imbalanced.

Report precision, recall, F1, sample size and the decision threshold together. For operational systems, consider the costs of different errors rather than choosing a model solely by its F1 score.

Where we use “F1 score”

AI Agents & AutomationMulti-Agent AI Failure Rate: What the 86.7% Figure Really Measures
Where does the 86.7% multi-agent failure rate come from, and what does it measure? A claim check of Berkeley, Google, MIT and Gartner sources shows why task structure matters more than any single rate.
Sep 13, 2026Key term

The longer version

Where we use “F1 score”

Get the Core in your inbox.