IIMT University Students: Choosing a Dataset for an AI Internship and Checking Its Licence

Choose a dataset that is small enough to finish in days, clean enough to model without weeks of repair, openly licensed for your intended use, and free of personal data. Read the licence text yourself and record it in writing before you write any code. In a 14-day internship, a wrong dataset choice on day 2 can cost you the whole project.

Why the dataset decision matters more in a 14-day AI internship

A short internship rewards students who reach a finished, explainable result. Most failed intern projects do not fail because of the model. They fail because the data was too large to download on a hostel connection, full of missing labels, locked behind a licence that forbids sharing, or contained personal details that should never have been posted on a public repository.

The free two-week online internship from CareerSwami gives you a 14-day window, a free offer letter and a free completion certificate after successful completion, across 26 tracks. Whatever track you take, the time limit is the same, so the vetting below is designed to take about half a day, not a week.

The four-part dataset vetting checklist: size, quality, licence, privacy

Run every candidate dataset through these four gates in order. Stop at the first gate it fails. Size and quality decide whether you can finish; licence and privacy decide whether you are allowed to use and show the result.

Gate 1: Size and format

• Download size: Prefer under 500 MB for a 14-day project. Anything in multiple gigabytes eats hours of download and processing time.
• Row count: For a tabular classification task, a few thousand to a few hundred thousand rows is workable on a laptop or free notebook service. Under a few hundred rows rarely supports a trustworthy model.
• Format: CSV, JSON, Parquet or a standard image folder is easiest. Proprietary or scanned-PDF-only data adds extraction work.
• Hardware fit: If it needs a GPU you do not have, pick a smaller subset or a different dataset.

Gate 2: Quality and labels

• Documentation: Is there a data card, README or column dictionary? No documentation usually means hidden surprises.
• Missing values: Open the file and count them per column. Above roughly 30 percent missing in a key column, expect real repair work.
• Label reliability: Who labelled it, and how? Crowd-labelled data can be noisy. Check a random sample of 20 rows by hand.
• Class balance: If 98 percent of rows belong to one class, accuracy will mislead you.
• Duplicates and leakage: Look for repeated rows and for columns that reveal the answer (for example, a “status” column that is derived from the target).
• Age and relevance: A 2012 dataset may not describe present-day behaviour.

Gate 3: Licence

The licence is the legal permission attached to the data. Absence of a licence does not mean “free to use”; it usually means no permission has been granted. The sections below show how to read it.

Gate 4: Privacy and sensitivity

• Does the data contain names, phone numbers, emails, addresses, government ID numbers, faces, voice or precise location?
• Does it contain health, financial, caste, religion or other sensitive categories?
• Does it involve minors?
• Was it scraped from social media or a website whose terms prohibit scraping?

If the answer to any is yes, choose another dataset unless the data is clearly anonymised and the publisher states that reuse is allowed.

Reading a dataset licence: what common licences allow in practice

These are general descriptions of widely used licences. Always open the actual licence text on the dataset page, because publishers sometimes modify standard terms.

Two further points are easy to miss. First, the code licence and the data licence on a repository can differ; a repository labelled MIT may still contain data with different terms. Second, a dataset hosted on a popular platform is not automatically cleared for reuse; the uploader may not own it.

Your licence log: a template to fill in on day 2

Keep a one-page record. If a mentor, a faculty member or a future employer asks where your data came from, you can answer in seconds.

1. Dataset name and version or download date
2. Source URL (the page, not just a direct file link)
3. Publisher or owner
4. Licence name and the URL of the licence text
5. Permitted uses you rely on (learning, non-commercial, research, commercial)
6. Conditions (attribution wording, share-alike, no redistribution)
7. Personal or sensitive data present? (yes/no, and what)
8. Whether you will publish the data itself, or only code and results
9. Your decision and the date

Scorecard for comparing three candidate datasets

Score each criterion from 0 to 2 (0 = fails, 1 = acceptable with work, 2 = clearly good). Any 0 in licence or privacy is a veto, whatever the total.

A total of 11 or more out of 14, with no vetoes, is a sound choice for a 14-day project.

Worked example: vetting three candidate datasets for a customer-churn project

This example is illustrative. The datasets and figures are invented to show the method; no code was run and no real dataset is being described.

A student wants to build a simple churn classifier. She shortlists three candidates.

• Dataset A: 7,000 rows, 2 MB CSV, a column dictionary, CC0 dedication, customer IDs only (no names). Missing values in one column (about 1 percent).
• Dataset B: 4 GB of call-detail records, no licence stated on the page, phone numbers included.
• Dataset C: 60,000 rows, CC BY-NC, partial documentation, 25 percent missing in the target label.

Scoring: A earns 2 on size, 2 on rows, 2 on documentation, 2 on labels, 2 on licence, 2 on privacy and 2 on fit, for 14. B scores 0 on size, 0 on licence and 0 on privacy, so it is vetoed immediately. C scores 2 on size, 2 on rows, 1 on documentation, 0 on labels, 1 on licence (non-commercial only), 2 on privacy and 1 on fit, for 9; its label problem alone would consume days.

She picks A, writes the log entry on day 2, and spends days 3 to 10 on cleaning, modelling and evaluation. Her report states the licence and the date accessed. Had she started with B, she would have discovered the problems only after a long download.

Privacy checks that apply to Indian student projects

India’s Digital Personal Data Protection Act, 2023 governs the processing of digital personal data. This article is not legal advice, and you should read the Act’s current text and any rules made under it if your project touches personal data. As a practical rule for interns, avoid datasets where individuals can be identified, never upload raw personal data to a public repository, and do not try to re-identify anonymised records.

• Do not collect your own data from classmates or the public unless you have explicit consent and a clear purpose.
• Remove direct identifiers (names, phone numbers, emails) before sharing even a sample.
• Beware of combined fields. Date of birth, pincode and gender together can identify a person.
• Prefer synthetic or aggregated data when your goal is only to demonstrate a method.
• Respect website terms. Scraping a site that forbids it can create problems even if the page is publicly visible.

Dataset checks for IIMT University students: tying your project to university rules

IIMT University is in Meerut, Uttar Pradesh, and its published Ordinances document gives the address as Ganga Nagar, Meerut-250001. The same document says the rules are issued by the university and that admission rules follow its Academic Council, which suggests regulations are set university-wide. It does not say what kind of university it is, so do not assume anything about its affiliation status.

For a data-focused internship, the relevant point is how project work is assessed. The M.Sc. (CS) ordinance says marks for sessionals, projects and industrial training follow the course structure prescribed by the university, and practicals, viva-voce, projects and reports are assessed by internal and external examiners. If your AI project report is ever examined, an external examiner may ask where your data came from and whether you were allowed to use it. Your licence log answers that directly.

The same ordinances also show an “Internship” paper (code 605, 100 marks) in semester 6 of the BA Journalism and Mass Communication scheme and a “Summer Training Project Report” worth 100 marks in the MBA semester 3 scheme. The extracted text we relied on does not give internship rules for B.Tech, BCA or MCA, and nothing in it covers online internships, credits or required documents. Whether a free online internship can count toward any of these papers is something only your department can decide. Treat this internship as extra experience unless the university confirms otherwise in writing, and do not assume it carries academic credit.

Verification steps:

1. Ask your department coordinator in writing whether online internships are accepted for your programme, and what document format is required.
2. Ask which examiner or faculty member will review project reports, and whether they expect a data-source declaration.
3. Request the current scheme or syllabus PDF for your course and look for the internship or project paper code.
4. Check whether the university has a policy on using personal or institutional data (for example, attendance or placement records) in student projects, and ask who approves such use.
5. Confirm with the training and placement office, if one exists for your programme, how an offer letter and completion certificate should be submitted.
6. Keep copies of every reply with dates alongside your licence log.

A 14-day plan that builds dataset vetting into the schedule

Warning signs that a dataset will cost you the internship

• The page offers a download but no named owner or publisher.
• The licence says “all rights reserved” or is missing.
• The file was uploaded by an individual with no explanation of how it was collected.
• Column names are cryptic (V1, V2, V3) with no dictionary, unless it is a well-known benchmark.
• The model scores near-perfect accuracy on day one, which often signals leakage.
• The data includes real names, phone numbers or faces and you have no consent trail.

Frequently Asked Questions

Can I use a Kaggle dataset for my AI internship project?

Often yes, but check the licence shown on that dataset’s own page. Kaggle hosts data under many different licences, and some uploads have unclear ownership. Record the licence in your log and cite the source.

What if a dataset has no licence at all?

Treat it as not cleared for reuse. Look for the original publisher, ask in writing for permission, or choose another dataset. Using it privately to learn is lower risk than publishing it, but a clearly licensed alternative is better.

Is a non-commercial licence a problem for a student portfolio?

Usually not for learning and a non-commercial portfolio, but it blocks use in a paid product or company work. Note the restriction in your report so nobody reuses your project commercially by mistake.

How many rows do I need for a meaningful result?

There is no fixed number. As a practical guide, you want enough examples in each class to train and hold out a test set. A few thousand rows is comfortable for a simple tabular task; very small datasets need careful cross-validation and cautious claims.

Can I upload the dataset to my GitHub repository?

Only if the licence permits redistribution and the data contains no personal information. Otherwise, publish your code and a link to the original source, and tell readers how to download it themselves.

Is anonymised data automatically safe to use?

No. Combinations of fields can sometimes identify people, and the publisher’s terms may still restrict reuse. Check the terms and never attempt to re-identify individuals.

Should I mention the dataset licence in my internship report?

Yes. Include the dataset name, source link, licence, access date and any attribution the licence requires. It shows professional habits and answers examiner questions before they are asked.

How does a dataset choice fit a free two-week internship like CareerSwami’s?

With only 14 days, pick a small, clean, openly licensed dataset by day 2. That leaves most of the time for modelling and writing, which is what your final submission and completion certificate depend on.

Related Artificial Intelligence guides

• Manipal University Jaipur Students: Cracking the AI Internship Behavioural Interview with STAR Stories
• Dr. Homi Bhabha State University Students: Applying AI to Science Data in a Short Internship
• M.V.N. University Students: Evaluating an AI Internship Certificate's Worth Beyond the Paper
• All Artificial Intelligence internship guides