Read the number
F1 ranges from 0 to 1 when defined in the usual way. It is high only when both precision and recall are high; a very low value for either pulls the harmonic mean down. F1 ignores true negatives, so it can hide performance that accuracy or calibration would reveal.
This page does not broaden into the excluded general precision vs recall curriculum. For the output in front of you, record the positive class, threshold, confusion counts, and averaging method.
| Average | What is combined |
|---|---|
| Binary | One designated positive class |
| Macro | Unweighted mean across class-specific F1 values |
| Weighted | Class-specific F1 weighted by support |
| Micro | Contributions pooled before the metric is computed |
The same predictions can receive different values under these averages.2
A small confusion-count example
Suppose the positive class has 30 true positives, 10 false positives, and 20 false negatives. Precision is 30/40 = .75 and recall is 30/50 = .60, so F1 is about .67. Changing the threshold can raise recall while lowering precision, or the reverse. The F1 value therefore belongs to that threshold; it is not a fixed property of the scoring model.
Now suppose the negative class is much larger. Accuracy could look high even while the positive-class F1 is weak. Conversely, a model optimized for positive-class F1 may produce too many false positives for a particular decision. Inspect the confusion counts rather than allowing the harmonic mean to erase the consequences.
Compare like with like
Use the same split, labels, threshold-selection process, and averaging definition. If the threshold was tuned on the test set, the reported F1 is optimistically biased. Repeated validation or an untouched test set is needed to separate tuning from final evaluation.
Compare F1 with a relevant baseline and with task-specific error costs. AUC interpretation covers threshold-spanning ranking; Brier score and calibration covers probability quality. Provenance rules explain why an unlabeled “.8 is good” table is not supplied.
A defensible reporting sentence
“Macro F1 was .74 on the held-out sample. Class-specific values ranged from .51 to .88, so the macro average did not hide the weaker class; threshold and prevalence were reported.”