Data Science Internship Viva Question Bank for KRMU

November 13, 2026

Web Stories

A data science internship viva usually tests whether you can explain what you did, why you chose it, what the data could and could not support, and where your results might mislead. This question bank groups likely questions under project, method, data and ethics, gives a model answer outline for each, and ends with a scoring sheet your peers can use in a mock viva.

How a Data Science Internship Viva Usually Works

An internship viva is a short oral examination, often alongside a report or presentation. The examiner reads your report, picks a few claims, and asks you to defend them. The format differs by institution and evaluator, so treat the questions below as a preparation pool rather than a guaranteed list.

Examiners tend to probe four things:

Ownership: which parts of the work were actually yours.
Reasoning: why you picked a method over alternatives.
Data awareness: where the data came from and what is wrong with it.
Responsibility: privacy, bias and honest reporting of limits.

For a 14-day internship, such as the free two-week online programme at CareerSwami, the project is small. That helps you: a compact project can be understood completely, and examiners respect a student who knows every line of a small piece of work more than one who gestures at a large one.

Preparing for a Viva After a 14-Day Data Science Internship

Before using the question bank, build a one-page project sheet. Every answer you give will draw from it.

One sentence stating the problem and who would use the answer.
The dataset: source, number of rows and columns, target variable, time period.
Cleaning steps you took, in order, with the reason for each.
The methods you tried and the metric you used to compare them.
Your final result, with its limits stated honestly.
Two things you would do with more time.

Spend the 14 days in a way that feeds this sheet. A workable split: days 1-2 on problem and data understanding, days 3-5 on cleaning and exploration, days 6-9 on modelling, days 10-11 on evaluation and error analysis, days 12-13 on the report, and day 14 on a mock viva.

Free 2-Week CareerSwami Internship

Online, 14 days, with a free offer letter and a free completion certificate once you finish the work. Check with your college that it fits your requirement before you enrol.

Apply for the free internship

Project Questions: Explaining What You Built and Your Role

These open almost every viva. Weak answers here make examiners dig harder everywhere else.

Likely question	Model answer outline
Summarise your project in one minute.	Problem, data, method, result, limitation, in that order. Use one number for the result and name the metric.
What was your own contribution?	List specific tasks you did yourself. If you used a tutorial or template, say so and state what you changed.
Why did you choose this problem?	Link it to a real decision someone could make. Avoid “it was assigned” as your only reason.
What was the hardest part?	Name one concrete obstacle, such as missing values or class imbalance, and how you resolved it.
How would you know the project succeeded?	State a success criterion defined before modelling, not after seeing results.
What would you change with another two weeks?	Give two realistic steps: more data, better features, error analysis, a simple deployment demo.
Can you walk me through your notebook or code?	Know the order of cells and what each does. Be ready to explain any line the examiner points to.

Method Questions: Defending Your Models and Metrics

Examiners use these to separate understanding from copying. Answer in plain language first, then add the technical term.

Likely question	Model answer outline
Why this algorithm and not another?	Start from the data: size, feature types, need for interpretability. Mention the baseline you compared against.
What was your baseline?	A simple reference such as predicting the majority class or the mean. Say how much your model improved on it.
Why did you choose that evaluation metric?	Tie it to the cost of errors. Accuracy misleads on imbalanced data; precision, recall or F1 may fit better.
How did you split your data?	Train/validation/test or cross-validation, with the reason. For time-ordered data, split by time.
What is overfitting and how did you check for it?	Good training score, poor unseen score. Describe the gap you measured and any step taken, such as regularisation or simpler features.
Which features mattered most?	Cite feature importance or coefficients, then caution that importance is not causation.
How did you tune hyperparameters?	Say what you tuned, the search method, and that tuning used validation data, not the test set.
Where does your model fail?	Show a few wrong predictions and a pattern in them. This single answer impresses more than a high score.

Method questions you can answer in two sentences

Bias versus variance: bias is error from overly simple assumptions; variance is sensitivity to the particular training sample.
Precision versus recall: precision asks how many flagged cases were right; recall asks how many real cases were caught.
Why scale features: distance-based and gradient-based methods are affected by feature ranges.
Why a test set: it gives an honest estimate of performance on unseen data, so it must stay untouched until the end.

Data Questions: Source, Quality, Cleaning and Leakage

Likely question	Model answer outline
Where did the data come from?	Name the source, licence or terms of use, and collection period. If it is a public dataset, say so.
How did you handle missing values?	State the share missing per column, the method (drop, impute, flag) and why it suits that column.
How did you deal with outliers?	Distinguish errors from genuine extremes. Explain what you kept and why.
Is your data representative?	Describe who or what is missing from it and how that limits where the result applies.
What is data leakage and did you avoid it?	Information from outside the training period or from the target slipping into features. Explain that you fitted scalers and imputers on training data only.
How did you treat class imbalance?	Report class proportions, then resampling, class weights or a different metric, with the effect observed.
How did you encode categorical variables?	One-hot, ordinal or target encoding, with the reason and any high-cardinality problem.
Can someone reproduce your results?	Mention fixed random seeds, saved cleaning steps and a recorded library version list.

Ethics Questions: Privacy, Bias and Honest Reporting

Many students skip this group, which is why a good answer stands out.

Does your data contain personal information? Say what identifiers exist, whether you removed or masked them, and that you would not publish raw personal records.
Could your model treat groups unfairly? Describe how you would compare error rates across groups such as gender, region or income band, if such fields exist.
Who could be harmed by a wrong prediction? Name the user, the type of error that hurts them most, and the human check you would keep in place.
Did you have permission to use the data? Refer to the licence or terms. If unsure, say so and say what you would check.
How would you communicate uncertainty? Report ranges or limits, not a single confident figure, and say what the model should not be used for.
Did you use AI tools or code from others? Disclose it plainly and show that you understand the result.

A Worked Illustration: Answering a Hard Follow-Up

Illustrative example only; the project and figures are invented for practice and no code was run.

Suppose a student built a model to predict whether a customer will cancel a subscription, using a public sample dataset. The examiner says: “Your accuracy is 91%. Is that good?”

A weak answer: “Yes, it is high, so the model is good.”

A strong answer, following the outline:

Challenge the number: “It depends on class balance. If 88% of customers do not cancel, predicting ‘no cancel’ for everyone already scores 88%.”
Give the better metric: “So I also checked recall on the cancel class, because missing a likely canceller is the costly error.”
Admit the limit: “Recall was modest, so the model is useful for ranking customers by risk, not for firm decisions.”
Offer a next step: “I would try class weights and look at the misclassified cancellers to find missing features.”

The pattern is: question the metric, show the better one, state the limit, propose a fix. It works for most follow-ups.

Mock Viva Scoring Sheet for Peer Panels

Run a 10-minute mock viva with two classmates as examiners. Each scores independently, then you compare notes. Score each row from 0 to 4 (0 = no answer, 2 = partly correct, 4 = clear and defended).

Criterion	What a 4 looks like	Score (0-4)
Project clarity	One-minute summary with problem, data, method, result, limit
Ownership of work	Names own tasks and what was borrowed
Method justification	Explains choice against a baseline and alternatives
Metric understanding	Links the metric to the cost of errors
Data handling	Explains source, missing values, leakage checks
Limitations	Shows specific failure cases without being asked
Ethics awareness	Addresses privacy, bias and permitted use
Communication	Plain language, steady pace, direct answers
Handling unknowns	Says “I am not sure, here is how I would check”

Reading the total out of 36

29-36: ready; practise timing.
21-28: solid base; repeat the weakest two criteria.
13-20: rebuild the project sheet and re-run the mock after a day.
Below 13: revisit your report line by line before the next attempt.

Have each examiner write one sentence of feedback per low score, not just a number.

Viva Preparation for K.R. Mangalam University Students: What to Check About Internship Evaluation

The university-wide calendar sits in the Student Handbook 2026-27. It lists a “Summer Internship Evaluation for Senior Batch” on 16-18 September 2026, and an “Internship” slot from 31 May 2027 onwards as per curriculum requirement. The handbook says the calendar can be revised, so check the current version before planning your preparation.

The B.Tech CSE programme handbook for 2026-30 describes 2-credit summer internships lasting 6 weeks, and an eighth-semester industrial or R&D project. It also mentions electives such as Natural Language Processing and Generative AI, which are useful topics to be able to discuss if your project touches them. A university blog post says the Career Development Centre handles internships and that students of all schools take part in summer internships as part of the curriculum.

What this means in practice: the university’s own internship evaluation is tied to its programme rules, which can differ by school. A short online internship may be a useful addition to your portfolio and a source of viva practice, but whether it satisfies any curriculum requirement, how the viva is marked, and what documents are needed should be confirmed with your school, not assumed. This article does not claim that a CareerSwami internship carries academic credit.

Verification steps before you prepare

Ask your programme coordinator in writing which semester your internship evaluation falls in and whether the viva is internal, external or both.
Request the evaluation scheme for your batch: marks split between report, viva and supervisor feedback.
Ask what documents must be submitted (certificate, report, any permission letter) and in what format.
Ask whether an online internship of two weeks is acceptable as additional work, or whether the curriculum requires the 6-week format.
Check the current academic calendar for revised evaluation dates.
Visit the Career Development Centre and ask for the internship guidelines that apply to your school.

Final Week Routine for the Internship Viva

Re-run your notebook from top to bottom to confirm it still works.
Rehearse the one-minute summary until it fits in 60 seconds.
Write ten questions about your own project that you hope nobody asks, then answer them.
Print your key plots and confusion matrix so you can point to them.
Do two peer mock vivas using the scoring sheet, a day apart.

Frequently Asked Questions

How long should my answers be in a data science internship viva?

Aim for 30 to 60 seconds per answer: the direct answer first, one reason, one example from your project. Let the examiner ask for more depth.

What if I do not know the answer to a technical question?

Say so, then explain how you would find out or what you do know nearby. Guessing confidently and being wrong costs more than honest reasoning.

Should I memorise formulas for the viva?

Know the ones tied to your project, such as the metric you reported, and be able to explain them in words. Pure formula recall matters less than interpretation.

How do I explain a result that was poor?

Report it plainly, show error analysis, and propose causes and fixes. Examiners often value a well-analysed weak result over an unexplained strong one.

Can I use pre-built libraries without being penalised?

Using libraries is normal. Be ready to explain what the function does, its key parameters and why you chose them.

How many of these questions should I rehearse?

Prepare all four groups, but give extra time to the questions that touch your actual project choices: your metric, your cleaning steps and your failure cases.

Does my viva format at K.R. Mangalam University depend on my school?

It may, since programme handbooks are school-level documents. Ask your coordinator for the written evaluation scheme for your batch rather than relying on seniors’ accounts.

How can a short two-week internship project still support a strong viva?

A small, fully understood project lets you answer follow-ups from first-hand knowledge. Keep scope narrow, document each decision, and practise explaining it aloud.