CBSE Class 10 · Artificial Intelligence · board exam 2027
Chapter 3: Evaluating Models — notes
The chapter’s 8 parts in the book’s order — the first idea of each and the points Medha Ma’am writes on the board — and the summary at the end. The full lesson and a check question for each part are in the app.
1. Why we evaluate an AI model
You already know four stages of the AI project cycle: Problem scoping, Data acquisition, Data exploration and Modelling.
- • Model evaluation = using metrics to understand a model's performance
- • Like a report card for your AI model
- • Build -> get feedback -> improve -> repeat
- • Goal: trustworthy and reliable AI systems
2. Train-test split
Okay, so, how do we check if a model is actually good?
- • Dataset -> training dataset + testing dataset
- • Training dataset: model learns
- • Testing dataset: predictions vs expected values
- • Aim: performance on new, unseen data
- • Testing on training data -> overfitting
3. Accuracy and Error
Okay, so let's talk about Bob and Billy.
- • Error = Abs (Actual-Predicted)
- • Error Rate = Error/Actual
- • Accuracy = 1-Error rate
- • Accuracy% = (Accuracy*100)%
- • 391k vs 402k: 11k, 0.027, 0.973, 97.3%
- • Goal: minimize error, maximize accuracy
4. Classification and the Confusion Matrix
Okay, so think about sorting supermarket items into two trolleys — one for fruits and vegetables, one for things like bread, oil and egg.
- • Confusion matrix: actual (y-axis) vs predicted (x-axis)
- • TP: positive correctly predicted
- • TN: negative correctly predicted
- • FP: negative predicted as positive
- • FN: positive predicted as negative
- • Disease activity: 07 correct out of 10
5. Classification accuracy and its limits
Okay, so — classification accuracy is just the number of correct predictions, out of all predictions.
- • Classification accuracy = (TP+TN)/(TP+TN+FP+FN)
- • Faulty model: TP=900, TN=0, FP=100, FN=0
- • 900/(900+0+100+0) = 0.9 = 90%
- • Accuracy suits only balanced datasets
- • Unbalanced: use Precision, Recall or F1 score
6. Precision
Okay, so let's talk about precision.
- • Precision = TP/(TP+FP)
- • Use when FPs must be reduced
- • Example: satellite launch weather
- • Spam: TP=240, FP=60
- • 240/(240+60) = 0.8 = 80%
7. Recall
Okay, so let's look at recall now.
- • Recall = TP/(TP+FN)
- • Also called Sensitivity or True Positive Rate
- • Use when FNs must be reduced
- • Examples: covid-19, fraud detection
- • Fraud: 80/(80+40) = 0.67
8. F1 Score
Okay, so — sometimes you just can't decide which matters more, FP or FN.
- • F1 Score = 2 x Precision x Recall / (Precision + Recall)
- • Use when unsure whether FP or FN matters more
- • Spam: TP=150, FP=50, FN=50
- • Precision = 0.75, Recall = 0.75
- • F1 = 1.125/1.5 = 0.75
Summary
Okay, so let's quickly look back at everything we covered today. We said model evaluation is like a report card for your AI model. And it gets better with good, honest feedback. Then we looked at the train-test split — how it divides data into a training dataset and a testing dataset. Do you remember why we never test on the training data? Right, because that causes overfitting. We also calculated three things — Error, Error Rate, and Accuracy. And the goal, always, is to minimize error and maximize accuracy. After that, we built a confusion matrix, with TP, TN, FP and FN. We saw something interesting there too — accuracy alone can mislead us, especially when the data is unbalanced. Finally, we learned three more ideas. Precision, for reducing False Positives. Recall, for reducing False Negatives. And the F1 Score, which balances both. So, out of everything we covered today, which part felt the most useful to you?
Ask Medha · Medha Ma’am — CBSE Class 10 Artificial Intelligence, Chapter 3. Written from the textbook, checked twice.
Questions parents and students ask
- How do I get the "Evaluating Models" notes as a PDF?
- Press "Save as PDF" — the phone’s print sheet opens; choose "Save as PDF" there. The PDF holds the notes only.
- Where are these notes written from?
- From the chapter’s own textbook text, part by part, in the book’s own terms; a second, independent model then checked every part against the book. The full lesson, a check question for each part and the practice are in the Ask Medha app.
The same preparation, with Medha Ma’am — on your phone
Every question of every chapter with its model answer, practice questions in the board’s own shapes, 45 minutes of class a day, written answers marked, a test every Sunday and a WhatsApp report to parents.
Medha Ma’am is an AI teacher, not a live tutor. Every answer is checked twice; a person takes responsibility.