📖 Chapter 6 of 13 · AI for New Learner- Senior 2 · Class 10
Chapter 6: Evaluation — Does It Actually Work?
In this chapter
- Mira from Chandigarh catches a data leakage bug.
- Why 99% accuracy can be 0% useful.
- Precision, recall, F1 — when each one matters.
- The confusion matrix — every ML engineer’s best friend.
- Real failure case: an AI loan model that rejected the wrong people.
- How to write a Model Card (what professionals do).
The Big Idea
Anyone can train a model. The hard skill is knowing whether it WORKS — and where it FAILS. A cancer-detection AI with 99% accuracy is useless if it misses 1 in 4 actual cancers. This is where junior engineers separate from senior ones. By the end of this chapter, you will evaluate models the way a senior engineer at Niramai or NPCI does.
Did you know?
In 2016, an AI screening tool used by a US hospital was 99% accurate at detecting one disease — but missed 50% of cases in dark-skinned patients. The training data was biased. Evaluation by sub-group is now a legal requirement in many AI deployments.
Builder Profile: Mira
Chandigarh · age 15
Mira interned with a Chandigarh startup that built a heart-attack-risk AI for clinics. The model showed 96% accuracy on internal tests. The founder was ecstatic. Mira looked closer.
She noticed something odd: the model used ‘time of day patient arrived’ as a feature. In the training data, severe cases came in at night. The model had learned: “midnight admission = high risk”. That is DATA LEAKAGE — the model was reading the answer indirectly from a feature that shouldn’t matter.
She rebuilt the evaluation. Real accuracy on fresh patients: 71%. The startup almost shipped a dangerous tool. Mira’s catch saved the company from a public scandal. Now she has a job offer for after her boards — at a Bengaluru AI safety company.
The confusion matrix
Every metric you need is hidden in one 2×2 grid. For a spam classifier, the four cells are: True Positive (spam correctly caught), False Negative (spam missed — it slipped through), False Positive (real mail wrongly blocked), and True Negative (real mail correctly delivered). Reading this grid is how you spot exactly where a model fails.
⊕ MATH MINUTE Precision, Recall, F1
PRECISION: of all things I called spam, how many actually were? Use when FALSE POSITIVES hurt (blocking real mail). RECALL: of all real spam, how many did I catch? Use when FALSE NEGATIVES hurt (cancer detection). F1: the harmonic mean of both — when you care about both equally.
Precision = TP/(TP+FP) • Recall = TP/(TP+FN) • F1 = 2PR/(P+R)
💻 CODE LAB Full Evaluation in One Block
from sklearn.metrics import (accuracy_score, precision_score,
recall_score, f1_score, confusion_matrix, classification_report)
# y_true: real labels y_pred: model predictions
print('Accuracy: ', accuracy_score(y_true, y_pred))
print('Precision:', precision_score(y_true, y_pred))
print('Recall: ', recall_score(y_true, y_pred))
print('F1: ', f1_score(y_true, y_pred))
print('Confusion matrix:')
print(confusion_matrix(y_true, y_pred))
# The professional shortcut:
print(classification_report(y_true, y_pred))
# Prints precision, recall, f1 for every class at once
⚠ FAILURE ANALYSIS
Problem: 99% accuracy but a useless cancer screener.
Why: Out of 1000 patients, 990 are healthy and 10 have cancer. The model predicts ‘healthy’ for EVERYONE. Accuracy = 990/1000 = 99%. But it caught 0/10 actual cancers — recall = 0%.
Fix: Use the right metric for the problem. For rare-event detection (cancer, fraud), RECALL matters more than accuracy. Also use stratified sampling, class weights, and the F1-score.
Solved illustrations
1. Why is accuracy a bad metric for cancer detection?
Show answer
Answer: If only 1% of patients have cancer, a model predicting ‘no cancer’ for everyone gets 99% accuracy but catches ZERO real cases. Use recall instead.
2. What is data leakage? Give one example.
Show answer
Answer: When information about the target variable accidentally appears in the features. Example: in Mira’s heart-risk model, ‘time of admission’ leaked the severity. Catching this matters more than the headline metric.
3. When is precision more important than recall?
Show answer
Answer: When false positives are costly. Spam filter: it is worse to block one important email (FP) than to miss one spam (FN). Use precision.
4. When is recall more important?
Show answer
Answer: When false negatives are costly. Cancer screening: missing one real cancer (FN) is far worse than scheduling one unnecessary follow-up (FP). Use recall.
5. What is a Model Card?
Show answer
Answer: A short document any ML team should write for every shipped model. It states: who the model is for, what data it was trained on, performance by sub-group, known failures, and last update date. Google, Hugging Face, and IndiaAI all require model cards.
Evaluation — your new habits
Before you ship ANY model, check all six: (1) train/test split done correctly, with the test set never seen during training and a fixed random seed; (2) the right metric for the problem — recall for cancer/fraud, precision for spam, F1 when both matter, not just accuracy; (3) evaluated by sub-group (gender, age, region), fixing any group that underperforms before shipping; (4) confusion matrix examined, with every false positive and negative understood; (5) failure cases documented; (6) a model card written with who built it, when, on what data, the performance numbers, and known limitations.
Questions — Chapter 6
- Write the formulas for precision and recall.
- Give one example of an Indian AI where high recall matters more than precision.
- What 4 things must every model card contain?
- Train any sklearn model and print its classification_report.
- Explain data leakage using Mira’s story in your own words.
Ship-It Challenge
Take any model from Ch 3-5. Write a 1-page Model Card. Share it on GitHub.
Check your understanding — Chapter 6
Tap an answer to check it instantly. These are for practice — nothing is saved or sent.
Q1. Why is accuracy a bad metric for cancer detection?
- Accuracy is always bad
- If only 1% of patients have cancer, a model predicting 'no cancer' for everyone gets 99% accuracy but catches ZERO real cases
- Cancer can't be detected by AI
Show answer
B. If only 1% of patients have cancer, a model predicting 'no cancer' for everyone gets 99% accuracy but catches ZERO real cases
Q2. What is data leakage?
- A broken hard drive
- When information about the target accidentally appears in the features — like 'time of admission' leaking severity in Mira's heart-risk model
- When data is deleted
Show answer
B. When information about the target accidentally appears in the features — like 'time of admission' leaking severity in Mira's heart-risk model
Q3. When is recall more important than precision?
- Never
- When false negatives are costly — e.g. missing a real cancer is far worse than one unnecessary follow-up
- Only for spam
Show answer
B. When false negatives are costly — e.g. missing a real cancer is far worse than one unnecessary follow-up
Q4. What is a Model Card?
- A playing card
- A short document stating who the model is for, its training data, performance by sub-group, known failures and last update
- A type of GPU
Show answer
B. A short document stating who the model is for, its training data, performance by sub-group, known failures and last update