F.S. University Students: Writing a Model Card for Your AI Internship Project

A model card is a short, structured document that says what your model is for, what data it learned from, how well it works and where it fails. For a 14-day AI internship, treat it as a standalone deliverable: someone who never opens your notebook should still understand and judge your work from the card alone.

This guide gives you a complete template, filled example sections for intended use, data and limitations, a worked illustrative card, a day-by-day plan and a review checklist. Everything is written for students at F.S. University in Shikohabad, but the template works for any AI project.

Why a model card deserves to be its own deliverable

Most student projects end with a notebook and a screenshot of an accuracy number. That tells a reviewer almost nothing. They cannot tell whether the accuracy came from a leaky split, whether the data represents real users, or whether the model should be used for anything beyond your demo.

A model card fixes this by forcing you to make claims you can defend. The idea was popularised by researchers who proposed short documents that accompany trained models and report performance across conditions. You do not need to follow any one published format. You need to be honest, specific and brief.

Treating it as standalone means three practical things:

• It is readable without running code. Anyone can understand it in five minutes.
• It is self-contained. Dataset name, version, split sizes and metric definitions are all written on the card.
• It is attachable. You can add it to a GitHub repository, a portfolio page or a resume link as a single PDF or Markdown file.

The model card template, section by section

Copy this structure. Each section has a purpose and a test for whether you have done it well.

Filled example: intended use

The following is illustrative and describes an imaginary project, not a real system.

Primary intended use: Classifying short English-language customer reviews of consumer electronics as positive, negative or neutral, as a learning demonstration of text classification.

Intended users: Students and reviewers evaluating the internship project.

Out of scope: Any decision about a real person or a real business. Not tested on Hindi, Hinglish or code-mixed text. Not suitable for sarcasm detection, medical text or legal text.

Notice the out-of-scope line. It is the most useful sentence on the card because it shows you understand where your work stops.

Filled example: data

Source: A public review dataset (name and version stated here), downloaded on a stated date.

Size: 12,000 reviews after removing duplicates and empty rows; 9,600 train, 1,200 validation, 1,200 test.

Labels: Derived from star ratings (1-2 negative, 3 neutral, 4-5 positive). These are proxy labels, not human-judged sentiment.

Split method: Random split with a fixed seed, after deduplication, so identical reviews do not appear in both train and test.

Preprocessing: Lowercasing, URL removal. No stemming.

The sentence about proxy labels matters. A star rating is not the same as sentiment, and saying so shows the reviewer you know the difference.

Filled example: limitations

The model confuses neutral reviews with positive ones more than any other pair (see the confusion matrix). Reviews under five words are often misclassified. All training text is English and from one product category, so performance on other categories is unknown. Star-rating labels contain noise, so the reported accuracy is an upper estimate of true sentiment accuracy. The model was not checked for bias against particular brands, names or dialects.

Each sentence names something checkable. Compare that with a weak version such as “The model may not work well in all cases”, which tells nobody anything.

Worked illustration: one complete card for a small project

Here is a compact version of a full card, again illustrative. The numbers are invented to show format and were not produced by running code.

Why macro F1 and not accuracy? If most reviews are positive, a model that always says “positive” looks accurate while being useless. Writing the reason for your metric on the card shows judgment, which reviewers value more than a high number.

A 14-day schedule for building the card alongside the project

A common error is to write the card on the last evening. Build it as you go, so each decision is recorded while you remember it. The free two-week online internship from CareerSwami runs 14 days, with a free offer letter and a free completion certificate after successful completion, so this plan fits that window.

Touch the test set only once, on day 10 or 11. If you tune against it, your reported number is optimistic and the card will be misleading.

Writing rules that make a card credible

• Use numbers with context. “0.71 macro F1 against a 0.28 baseline on 1,200 test reviews” beats “good performance”.
• Separate what you measured from what you believe. Label guesses as guesses.
• Name the failure, not the category. Show two or three real misclassified examples.
• State the label origin. Human-labelled, scraped, or proxy: say which.
• Keep it short. Two pages is enough for an internship project.
• Do not oversell. A card that admits weaknesses is more trusted than one that lists none.

Review checklist before you submit the card

1. Is the model version and date written on the first page?
2. Does the intended-use section include at least one out-of-scope use?
3. Can a stranger find the dataset, its version and your split sizes?
4. Is there a baseline, and is the metric choice explained?
5. Was the test set used only once for the final number?
6. Is at least one slice of results shown, or the absence explained?
7. Does the limitations section include real error examples?
8. Have you removed or avoided personal data, and does the card say so?
9. Can someone reproduce your result from the notes given?
10. Did someone outside your project read it and understand it?

F.S. University students: placing your model card within your programme’s project requirements

If you study at F.S. University, a model card is useful both as a portfolio item and as supporting material for any project or internship submission your department asks for. The B.Tech page lists specialisations including CSE, IT and B.Tech in AI & Data Science, so AI projects are a natural fit for many students. That page gives no syllabus or credit structure, so it does not tell you how project work is assessed.

The 2025-26 academic calendar carries university-wide dates for the Summer Project/Internship/Dissertation, including allotment, submission and evaluation windows. It sets no duration, credits or rules, and it covers 2025-26, so look for the current year’s version. Its even-semester start entries also appear to contain a year typo, so do not rely on them without checking.

The Training & Placement Cell page lists an “Internship Support” step and mentions “6-month guaranteed internships”. It gives no scheme year, credits, evaluation rules, documents or policy on online internships. So a short online internship is not confirmed to count for anything formally. Treat the card as proof of your work, and do not assume it satisfies a university requirement until your department says so.

Verification steps for F.S. University students before you rely on your card

1. Email your department coordinator and ask in writing whether an online internship of 14 days can be submitted for any project or internship requirement.
2. Ask which document sets your project rules, such as a regulation, scheme booklet or ordinance, and request the current year’s copy.
3. Ask whether evaluators expect a particular report format, and whether a model card can be attached as an annexure.
4. Check the current academic calendar for the project allotment, submission and evaluation dates that apply to your semester.
5. Ask the Training & Placement Cell what proof of internship they accept, for example a certificate or an offer letter.

Frequently Asked Questions

How long should a model card for a 14-day internship project be?

One to two pages is enough. Aim for dense, specific sentences in each of the eight sections rather than long prose. If a section runs past half a page, you are probably repeating your report.

Do I need a model card if my project is just a notebook?

Yes, and it matters more then. A notebook shows code but rarely states intended use, label origin or limitations. A card makes those visible and lets a reviewer judge the work without running anything.

What if my model performs poorly?

Report it honestly. A weak result with a clear baseline, an error analysis and a sensible explanation is a stronger deliverable than an inflated number you cannot defend. The limitations section is where that analysis belongs.

How do I handle subgroup evaluation with a small dataset?

Pick a slice you can actually compute, such as text length, image brightness or class. If a slice has very few examples, report the count and say the estimate is unreliable. Stating that you could not evaluate a slice is better than leaving it out silently.

Should the card include my training code?

No. Include reproduction notes: seed, library versions, the run command and approximate runtime. Link to the repository if you have one, but keep the card readable by itself.

Which file format should I submit the card in?

Markdown works well for GitHub, and PDF is safer for sharing with evaluators or recruiters. Keep both in sync, and put the version and date at the top so people know which one is current.

Can I reuse the same card template for later projects?

Yes. Keep the eight headings and rewrite every entry for the new model and data. Do not carry over limitations or ethical notes without checking them, because they are usually specific to the dataset.

How does a model card strengthen a completed CareerSwami internship project?

The internship gives you a certificate after successful completion, but the certificate does not describe your work. A model card is the document that does: it shows what you built, how you tested it and where it falls short, which you can attach to a portfolio or resume.

Related Artificial Intelligence guides

• ASBM University Students: AI for Business Analytics Internships, From Dashboard to Prediction
• Netaji Subhas University Students: Setting Up a Reproducible Python Environment for AI Internship Work
• Manipur International University Students: Getting Mentor Feedback Remotely on Your AI Work
• All Artificial Intelligence internship guides