Learn AI /Model Quality and Evaluation /F1 score
F1 score
- Deeper
- Model Quality and Evaluation
- Technical readers
- Programme & data teams
Also called: F-score, F1.
Back to the milestone you came from
In five seconds
What it is
A single score that balances how many positive cases a system finds with how often its positive predictions are correct.
In plain words
F1 score gives one number for how well a system finds positive cases without producing too many wrong positive alerts.
An everyday example
Amina is tuning a radio in Amman to hear a weather bulletin from a distant station. If she turns the dial too broadly, she hears many bursts of sound that are not the station she wants. If she tunes too narrowly, the sound is cleaner, but parts of the real broadcast fade out. A good setting needs both a clear signal and enough of the actual programme to come through. The F1 score is like giving one summary mark for that balance.
A radio dial is a continuous physical signal, while F1 is calculated from labelled examples after predictions are made. Real classification problems may involve thresholds, class imbalance and unequal harms, none of which the radio story shows well. Also, radio quality has many dimensions, but F1 only combines precision and recall.
Remember thisF1 summarises the balance between precision and recall, with the lower value pulling the score down. It is useful under class imbalance, but excludes true negatives and cannot show whether particular errors are acceptable.
Understand it
Understand it
Why it matters
A single score can make model results look simple, but decisions are rarely simple. F1 matters because it is often used when accuracy would be misleading, especially if positive cases are rare. For example, a system can look highly accurate by mostly saying “not positive”, yet still miss many of the cases people care about. F1 helps you ask a better question: does the system both find important cases and avoid too many false alarms? That said, it should not be the only basis for procurement, deployment or risk decisions. If missing a case is much more harmful than raising a false alarm, or the reverse, then two systems with the same F1 may still be very different in practice.
How F1 score is built from prediction counts
After viewing this, a reader should be able to say that F1 starts from true positives, false positives and false negatives, turns them into precision and recall, and then combines them so one weak part pulls the score down.
01
Count outcomes
Compare each prediction with the real label and count true positives, false positives and false negatives.
Draw two columns labelled 'Predictions + labels' and 'Counts'. In the counts area, show three stacked boxes: TP, FP, FN, with short captions beneath each.
02
Calculate precision
Precision uses true positives and false positives. It asks how reliable the positive predictions were.
Draw an arrow from TP and FP to a box labelled 'Precision = TP / (TP + FP)'. Use a brief note: 'fewer false alarms raises precision'.
03
Calculate recall
Recall uses true positives and false negatives. It asks how many real positive cases were found.
Draw an arrow from TP and FN to a box labelled 'Recall = TP / (TP + FN)'. Use a brief note: 'fewer misses raises recall'.
04
Combine into F1
F1 combines precision and recall with the harmonic mean, so one low value keeps the final score low.
Draw arrows from the precision and recall boxes into a final box labelled 'F1 = 2PR / (P + R)'. Add a small callout: 'high only when both are reasonably high'.
A left-to-right flow diagram. Start with model predictions compared with real labels. From that, show three counts: true positives, false positives and false negatives. These feed into two boxes, precision and recall. A final box combines them into F1, with a note that both need to be strong for F1 to be strong.
Hover or tap each box to reveal a one-line definition and a small numeric example showing how a low precision or low recall reduces the final F1.
Describe this without the diagram
This flow diagram explains the procedure for getting an F1 score. First, compare model predictions with real labels and count three outcomes: true positives, false positives and false negatives. Next, calculate precision from true positives divided by true positives plus false positives, which reflects how trustworthy positive predictions are. Then calculate recall from true positives divided by true positives plus false negatives, which reflects how many real positive cases were found. Finally, combine precision and recall using F1 equals two times precision times recall divided by precision plus recall. The key message is that the final score stays low if either precision or recall is low.
With reduced motion: A static left-to-right chart with four panels: counts, precision, recall and F1. Arrows show that TP, FP and FN lead to precision and recall, and those two lead to F1. A note under the last panel says the score rewards balance, not one strong value alone.
F1 score is used for tasks with two classes, such as “urgent” and “not urgent” or “fraud” and “not fraud”. It combines two other measures: precision and recall. Precision asks, “When the system says positive, how often is it right?” Recall asks, “Of all the real positive cases, how many did it catch?”
What goes in are counts from a confusion matrix: true positives, false positives and false negatives. True positives are positive cases found correctly. False positives are cases wrongly marked positive. False negatives are positive cases the system missed. From those counts, you calculate precision and recall, then combine them into one number.
What happens in the combination matters. F1 uses the harmonic mean, which gives a low result if either precision or recall is low. So a system cannot get a strong F1 score by doing very well on one and badly on the other. If precision is 0.90 and recall is 0.30, the F1 score stays much closer to the weaker value than to the stronger one.
When defined, F1 is between 0 and 1. Higher usually means a better balance between missed positives and false alarms. But it is not a complete picture. It does not include true negatives, so it can hide some kinds of failure. It also does not tell you whether predicted probabilities are well calibrated, and it does not tell you whether false positives and false negatives have equal real-world cost.
How “Tuning a radio for clear reception” maps onto the real thing
| In the story | In the AI system |
|---|---|
| clear parts of the station that are truly the right broadcast | correct positive predictions |
| static or other stations mistaken for the target signal | false positives |
| parts of the real broadcast lost because tuning is too narrow or weak | false negatives |
| one judgement of whether the tuning balances clarity and coverage | F1 score balancing precision and recall |
A worked example
Before the progress report enters clearance, the partnerships officer evaluates the classifier against needs assessment answers previously coded and checked by two reviewers.
Starting from: Among 200 checked answers, 40 require partner follow-up. The classifier flags 50 answers, of which 30 genuinely require follow-up.
- Count the outcomes: true positives = 30, false positives = 20 and false negatives = 10.
- Calculate precision: 30 / (30 + 20) = 0.60. Of the answers flagged for follow-up, 60 per cent were correctly flagged.
- Calculate recall: 30 / (30 + 10) = 0.75. The classifier found 75 per cent of all answers requiring follow-up.
- Combine them: F1 = 2 × 0.60 × 0.75 / (0.60 + 0.75) = 0.67, rounded.
Result: The F1 score is about 0.67. The classifier finds most follow-up items, but 20 of its 50 flags are false alarms and it misses 10 relevant answers. The officer should examine those errors and the operational threshold, not treat 0.67 as a verdict on whether the model is fit for use.
In a United Nations workplace
A partnerships officer in a country office is testing a classifier while coding open-ended answers from a needs assessment. The model tags answers that may require partner follow-up before the progress report goes for clearance at close of business.
F1 helps compare model versions when both missed follow-up items and unnecessary flags create work. It combines the share of flagged answers that reviewers confirm as relevant with the share of all relevant answers that the model finds. Because the lower value weighs heavily, strong recall cannot conceal weak precision, or vice versa.
Before you try thisNeeds assessment answers may contain personal or protection-sensitive details. Raw answers must not be entered into a tool that has not been approved for that material, because doing so could expose respondents. Use an approved environment, minimise the evaluation data and retain human review. Also inspect false positives and false negatives separately, since F1 excludes true negatives and does not reflect the different consequences of each error.
Where you may hear it
- A vendor demo during a procurement meeting
- An ICT governance board reviewing a classification tool
- Clearance of a model evaluation report
- A progress meeting with an implementing partner
Avoid confusion
Avoid confusion
Often confused with
Common misconceptions
Common misconception: A higher F1 score always means the better system.
Correction: Not always. F1 only balances precision and recall. It ignores true negatives and does not reflect whether false positives and false negatives have different real-world costs. In a medical, safety or humanitarian setting, the better system may be the one with lower F1 but fewer harmful misses, depending on the use case.
Common misconception: F1 is just another name for accuracy.
Correction: They measure different things. Accuracy counts all correct predictions, including true negatives. F1 does not use true negatives at all. When one class is much more common than the other, accuracy can look high even if the system performs poorly on the cases you care about, while F1 focuses on positive-case performance.
Common misconception: If precision and recall are both decent, the exact balance does not matter.
Correction: The balance does matter because F1 drops when one value is much lower than the other. A system with precision 0.90 and recall 0.50 does not get the same F1 as a system with precision 0.70 and recall 0.70. F1 rewards balance more than a simple average would.
Quick check
Why can F1 be more informative than accuracy when positive cases are rare?
When positive cases are rare, a model can get high accuracy just by mostly predicting the negative class. F1 helps counter that by combining precision and recall, which focus on finding real positives and limiting false positive alerts. It still does not replace looking at the underlying error counts.
Go deeper
Go deeper
The simple version
F1 score gives one number for how well a system finds positive cases without producing too many wrong positive alerts.
The technical version
For binary classification, F1 is the harmonic mean of precision and recall: F1 = 2PR / (P + R), where precision P = TP / (TP + FP) and recall R = TP / (TP + FN). Because it is a harmonic mean, the result is strongly influenced by the smaller of the two values, so it rewards balance rather than allowing one high value to hide one low value. F1 is especially common under class imbalance, where plain accuracy can be misleading. However, it excludes true negatives, so it should not be treated as a full description of performance. Variants such as macro, micro and weighted F-scores extend the idea to multi-class settings, and F-beta variants change the balance to favour recall or precision more strongly.
Reviewed external resources
Primary sourceAI Risk Management Framework (opens in a new tab)NIST
This link gives a broader framework for managing AI risk in practice, and suits readers who need to place metrics like F1 within governance, testing and human oversight rather than treating one score as enough.
Intermediatelink checked 2026-07-31
EncyclopaediaF-score (opens in a new tab)Wikipedia
This link gives a broad reference-style overview of the F-score family, including formulas and related variants, and suits readers who want a quick encyclopaedic summary.
Intermediatelink checked 2026-07-31
CourseClassification: Accuracy, recall, precision (opens in a new tab)Google
This link gives a practical teaching explanation of classification measures such as accuracy, recall and precision, and suits readers building basic evaluation literacy before interpreting F1.
Intermediatelink checked 2026-07-31
VideoVideo explanations: F1 score (opens in a new tab)YouTube search
This link is a YouTube search results page for video explanations of F1 score, so it suits readers who prefer to compare several visual teaching styles rather than relying on one source.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: F1 score (opens in a new tab)Reddit search
This link is a Reddit search results page for community discussion about F1 score, so it suits readers who want to see common points of confusion and informal explanations, but claims there should be checked carefully.
search results, not one pagelink checked 2026-07-31
Where this came from
Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.
The one thing to remember
F1 summarises the balance between precision and recall, with the lower value pulling the score down. It is useful under class imbalance, but excludes true negatives and cannot show whether particular errors are acceptable.