Learn · Try · Check
What you’ll learn
Choose measures that reveal the errors your task cares about.
Listen to this lesson
Local AI narration · Quiz answers are kept out of the audio.
Read the audio transcript
ML 101. Accuracy, precision and recall. Choose measures that reveal the errors your task cares about. Count outcomes before choosing a headline score. For a binary review task, call an example positive when it truly needs review. A true positive is correctly flagged. A false positive is routine material unnecessarily flagged. A false negative is a missed review case. A true negative is routine material correctly left unflagged. These names depend on which class you designate positive, so define that class explicitly. Work through one small example. Imagine 100 cases: 10 truly need review and 90 are routine. A teaching model flags 8 of the 10 review cases and also flags 4 routine cases. That gives 8 true positives, 2 false negatives, 4 false positives and 86 true negatives. Accuracy is (8 + 86) / 100 = 94%. Precision is 8 / (8 + 4), about 67%: of the flagged cases, how many truly needed review? Recall is 8 / (8 + 2) = 80%: of the true review cases, how many were found? These are illustrative numbers, not a HerbDev product benchmark. Connect the measure to the decision. A system that flags nothing would be 90% accurate on this dataset while missing every review case. Accuracy alone hides that failure. If missed cases matter most, examine recall and the missed examples. If the review queue is overloaded, inspect false positives and precision too. Changing a decision threshold can trade one error pattern for another. Choose the threshold using validation data and the workflow’s costs; report the resulting test performance honestly. HerbDev example. HerbDev’s comparison view distinguishes accuracy and precision. Return to it whenever a report uses one score to imply a different kind of reliability. Precision is not a synonym for overall correctness. The process: Define the positive class. Then Count TP / FP / FN / TN. Then Compute relevant measures. Then Inspect the error tradeoff. Try this. Using the 100-case example, explain why the model’s 94% accuracy does not mean that 94% of its flags are correct. Pause the audio to try the exercise. A worked answer and checkpoint questions are available on the lesson page.
Count outcomes before choosing a headline score
For a binary review task, call an example positive when it truly needs review. A true positive is correctly flagged. A false positive is routine material unnecessarily flagged. A false negative is a missed review case. A true negative is routine material correctly left unflagged. These names depend on which class you designate positive, so define that class explicitly.
Work through one small example
Imagine 100 cases: 10 truly need review and 90 are routine. A teaching model flags 8 of the 10 review cases and also flags 4 routine cases. That gives 8 true positives, 2 false negatives, 4 false positives and 86 true negatives. Accuracy is (8 + 86) / 100 = 94%. Precision is 8 / (8 + 4), about 67%: of the flagged cases, how many truly needed review? Recall is 8 / (8 + 2) = 80%: of the true review cases, how many were found? These are illustrative numbers, not a HerbDev product benchmark.
Connect the measure to the decision
A system that flags nothing would be 90% accurate on this dataset while missing every review case. Accuracy alone hides that failure. If missed cases matter most, examine recall and the missed examples. If the review queue is overloaded, inspect false positives and precision too. Changing a decision threshold can trade one error pattern for another. Choose the threshold using validation data and the workflow’s costs; report the resulting test performance honestly.
- Define the positive class
- Count TP / FP / FN / TN
- Compute relevant measures
- Inspect the error tradeoff
Try it yourself
Using the 100-case example, explain why the model’s 94% accuracy does not mean that 94% of its flags are correct.
Compare with a worked answer
Accuracy counts all correct decisions, including 86 routine cases. Only 8 of the 12 flags are true review cases, so the fraction of correct flags—precision—is about 67%.
Check your understanding
Choose one answer for each question, then check your score. Read the explanations and retry any question you missed.
Read the answer guide
1. 12 There are 8 true positives plus 4 false positives, for 12 flags.
2. 8 divided by 10, or 80% Recall measures the fraction of actual positives found: TP / (TP + FN).
3. It misses every case that needs review The majority class dominates accuracy. The baseline has zero recall for the review class.
Was this explanation clear?
Help improve this lesson. With cookies accepted, HerbDev records one current rating per lesson for this browser. Server checks limit repeated submissions; accepted changes may also appear in site analytics. No written answer or account is needed.
You can also share a suggestion with HerbDev.
Review missed questions
Revisit questions missed on your latest checkpoint attempts. These short reviews help you practise; they do not change your best checkpoint scores. Return to the lesson checkpoint when you are ready.
Check a lesson’s answers to build your review list. Saved reviews are available after cookie acceptance.
Course complete. Keep the questions coming.
You have answered every checkpoint correctly. Use your worked exercises to explain the ideas to someone else, then apply them to a small project.
Apply it in the final project · Choose what to learn next →Keep exploring: What next? · AI vocabulary · AI 101 · Learning hub