Final Book Exam

Total 60 marks. Structure: A. Fill in the blanks (10), B. Code reading — predict the output (15), C. Design questions (15), D. Long answer, pick ONE (20).

A. Fill in the blanks (10 marks)

1. The Python library used in this book to load and save trained models is .
2. The activation function that outputs values between 0 and 1 is the function.
3. RECALL = TP / (TP + ).
4. India’s data protection law passed in 2023 is the Act.
5. Hugging Face’s free deployment platform for ML web apps is called .
6. The ‘Hello World’ dataset of deep learning is .
7. The Streamlit command to run an app locally is: streamlit app.py.
8. India’s Govt open data portal URL is .
9. A document describing a model’s intended use, training data and performance is a Model .
10. If a model has 99% accuracy on imbalanced data, the metric to also check is .

B. Code reading — predict the output (15 marks)

1. import numpy as np; x = np.array([1,2,3,4]); print(x.mean() + x.std().round(2)) →
2. m = LinearRegression(); m.fit([[1],[2],[3]], [10,20,30]); print(m.predict([[5]])) →
3. TP,FP,FN = 80,20,10; P = TP/(TP+FP); R = TP/(TP+FN); print(round(2*P*R/(P+R), 2)) →
4. df = pd.DataFrame({'p':[80,70,90,60]}); print(df[df.p > 65].p.median()) →
5. m = DecisionTreeClassifier(max_depth=1); m.fit([[0],[1],[2],[3]], [0,0,1,1]); print(m.predict([[1.5]])) →

C. Design the system (15 marks)

1. You must build an AI that screens job applicants for an Indian IT company. List FIVE things you would do to ensure your model is fair and complies with the DPDP Act. (5 marks)
2. A school wants an AI that detects bullying from school WhatsApp messages. Identify TWO ethical problems, and propose how you would solve them — or refuse the project. (5 marks)
3. A startup wants to predict which farmers are at risk of crop failure using public data. Discuss why predicting on individuals can be unfair, AND what a responsible alternative might be (e.g. district-level advisory instead of individual prediction). (5 marks)

D. Long answer — pick ANY ONE (20 marks)

1. Describe the full ML lifecycle as you would apply it to building a Bhashini-style translator for any TWO Indian languages. Cover: data collection (with sources), preprocessing, model choice, training, evaluation (with metrics), deployment, model card, and DPDP compliance. Mention real Indian datasets and tools by name.
2. Pick any TWO chapters of this book. For each, identify the MOST important lesson, the builder you found most inspiring, and one project idea you have now that you didn’t have before. Explain WHY you would build it for India.
3. Compare scikit-learn, PyTorch and Streamlit. For each: (i) what problem it is best for, (ii) one Indian company/lab that uses it, (iii) one limitation, (iv) ONE project you would build with it. Write as a structured 400-word essay.