How to Build a Data Science Portfolio With Free Datasets in 2026

Here is something most data science students find out too late.

Recruiters do not hire certificates. They hire proof.

A certificate tells someone you sat through a course. A finished project tells them you can actually do the job. Three finished projects, each with clean code and a clear explanation of what you did and why, will get you further in interviews than ten half-done tutorials you never completed.

The good news: every dataset you need to build a genuinely impressive portfolio is completely free. The right dataset can make the difference between a project that showcases your skills and one that gets lost in data cleaning headaches.

This guide gives you the free datasets, the specific projects to build with them, and a clear sequence from beginner to job-ready. No paid courses. No premium dataset subscriptions.

What this guide covers

Why a portfolio matters more than a certificate
What makes a data science portfolio project actually impressive
The best free dataset sources in 2026
12 specific projects from beginner to advanced with dataset links
Pakistan-specific project ideas that stand out locally
How to present your portfolio on GitHub
Common mistakes that make portfolios forgettable
The realistic timeline to a job-ready portfolio

What makes a portfolio project actually impressive

Before choosing a dataset, understand what Pakistani and international employers look for when they review a data science portfolio.

Employers want to see your thinking process, not just polished results. Document your approach, explain your decisions, and share what you learned. That is what makes a portfolio stand out.

A project impresses when it has four things:

A clear problem statement. Not "I analysed the Titanic dataset." But "I built a model to predict passenger survival and identified that ticket class was a stronger predictor than age, which challenges the common assumption."

A documented decision process. Why did you choose this model over others? What did you try that did not work? What would you do differently with more data?

Reproducible code. Clean, commented Python or R in a well-structured GitHub repository. Anyone should be able to clone it and run it.

An insight that is actually useful. The best projects answer a question a real business or person would care about. Not just "the model achieved 87% accuracy" but "this model could reduce customer churn by identifying at-risk customers two months before they leave."

The best free dataset sources in 2026

These are the most reliable and well-documented free dataset platforms. Every dataset listed in this guide is available from these sources at no cost.

Kaggle (kaggle.com)

Kaggle is one of the most popular data science platforms. It hosts competitions and a catalogue of courses across machine learning and AI, and it offers thousands of datasets, big and small, that you can download for free.

Kaggle is your starting point. Create a free account, download datasets directly, and run notebooks in the browser without any local setup. For Pakistani students with limited laptop RAM, Kaggle notebooks run entirely in the cloud.

UCI Machine Learning Repository (archive.ics.uci.edu)

The oldest and most academically reliable dataset repository. Over 600 datasets spanning classification, regression, clustering, and time series. Less polished than Kaggle but more varied in problem types.

Google Dataset Search (datasetsearch.research.google.com)

Google's search engine specifically for datasets. Type any topic and it finds publicly available datasets from government agencies, academic institutions, and research organisations worldwide.

World Bank Open Data (data.worldbank.org)

One of the richest, most diverse resources of statistical facts and public datasets. You can search by category such as country or indicator, and DataBank helps you analyse and visualise large datasets.

For Pakistani students wanting projects with local economic relevance, the World Bank has extensive Pakistan data covering GDP, inflation, employment, education, and health indicators.

Pakistan Bureau of Statistics (pbs.gov.pk)

The official Pakistani government statistics portal. Census data, labour force surveys, household income and expenditure surveys, and economic indicators. Building a project with PBS data is a genuine differentiator in Pakistani job applications.

Stack Overflow Developer Survey (survey.stackoverflow.co)

A large, well-documented survey dataset that is excellent for segmentation, trend analysis, and communicating insights to a non-technical audience. The 2025 survey reports 49,000+ responses across 177 countries, covering demographics, work context, tools, databases, languages, and AI usage patterns.

Particularly relevant for Pakistani developers because you can filter for South Asian respondents or compare Pakistani developer salaries and tool usage against global benchmarks.

The 12 projects: beginner to advanced

These projects are organised in order of difficulty. Start at your level and build upward. Each one has a specific dataset, a clear problem, and the skills it demonstrates.

Beginner projects (month 1 to 2)

Project 1: Student performance analysis

Dataset: Student Performance Dataset on Kaggle (by Paulo Cortez), 649 students, 30 features. Skills: pandas, data cleaning, correlation analysis, matplotlib, data storytelling.

Load the dataset containing student grades and demographic information. Clean the data. Explore which factors correlate most strongly with final grade. Build a visualisation showing the relationship between study time, family support, and academic performance.

The insight to find: does study time actually predict grades? Most analyses find it does, but with diminishing returns above a certain threshold. Study habits, sleep, and screen time make this one of the most relatable beginner projects because you are basically analysing your own life with code.

Document which variables you expected to matter but did not, what surprised you, and what a school could practically do with this information.

Project 2: Pakistan economic indicators dashboard

Dataset: World Bank Open Data, Pakistan indicators (GDP, inflation, unemployment, literacy). Skills: pandas, cleaning, time series visualisation, Matplotlib or Plotly, storytelling.

Download 20 years of Pakistan economic data. Build an interactive dashboard showing trends in GDP growth, inflation, unemployment, and literacy. Highlight the 2022 to 2024 economic crisis and the recovery.

Using Pakistan-specific data with global context immediately differentiates you from candidates using generic datasets. Pakistani employers find it directly relevant, and international employers see your ability to work with real-world context.

Project 3: Mobile app review analysis

Dataset: Google Play Store dataset on Kaggle. Skills: pandas, cleaning, frequency analysis, basic visualisation, data quality handling.

Pick a category of apps and figure out what separates a 4.5-star app from a 2-star app. Clean messy review counts, handle duplicates, and build frequency tables that tell a story. For a Pakistani context, filter for apps popular here: food delivery, ride-hailing, mobile banking, education. What makes one fintech app rated 4.5 while a similar competitor sits at 2.8?

This project is primarily about data cleaning, which is 80% of real data science work. Showing you can handle messy real-world data is more impressive than a clean tutorial dataset.

Project 4: COVID-19 data visualisation

Dataset: Our World in Data COVID-19 dataset (free, regularly updated). Skills: pandas, time series analysis, multi-country comparison, Plotly or Seaborn.

Analyse trends, predict cases, or visualise recovery rates. Compare Pakistan's trajectory against comparable economies like Bangladesh, India, and Iran. How did Pakistan's vaccination rate compare regionally? How did economic indicators correlate with case severity?

COVID data is clean, well-documented, globally available, and tells a genuinely important story. The comparison angle adds analytical depth that pure trend analysis lacks.

Intermediate projects (month 2 to 4)

Project 5: Customer churn prediction

Dataset: Telco Customer Churn dataset on Kaggle (IBM sample data). Skills: scikit-learn, logistic regression, random forests, model evaluation, feature importance, class imbalance.

Build a model that predicts which telecom customers are likely to cancel. The dataset is imbalanced, so you need SMOTE or class weighting, which shows you understand real-world ML beyond clean textbook examples.

Frame it for business: "This model correctly identifies 78% of customers who will churn, allowing the company to target retention offers and potentially cut churn by 15%." Churn work teaches classification, imbalanced data, feature engineering, and business communication, all skills that transfer directly to Pakistani fintech, telecom, and e-commerce roles.

Project 6: House price prediction

Dataset: Ames Housing dataset on Kaggle (2,930 houses, 79 features). Skills: feature engineering, linear regression, gradient boosting, cross-validation, missing values.

Predict sale prices from property features. Do not just build a model, explain which features matter most. What is the price premium for an extra bathroom? How much does lot size matter versus neighbourhood?

For Pakistan context, extend the write-up by discussing how you would build a similar model for Lahore or Karachi property data. That shows contextual thinking.

Project 7: Sentiment analysis of Pakistani social media

Dataset: X/Twitter API free tier, or existing social media datasets on Kaggle. Skills: NLP, text preprocessing, sentiment classification, NLTK or spaCy.

Collect posts about a Pakistani topic, cricket, the budget announcement, the tech job market, and classify sentiment as positive, negative, or neutral. Visualise how sentiment shifted over time or across events.

NLP on Urdu-English mixed text is a real, unsolved problem in Pakistan. Even a basic analysis that acknowledges code-switching and attempts to handle it shows more sophistication than a pure English sentiment project. Document how you handled Roman Urdu, your preprocessing decisions, and your limitations.

Project 8: Financial fraud detection

Dataset: Credit Card Fraud Detection dataset on Kaggle (284,807 transactions). Skills: severely imbalanced data, precision-recall tradeoff, anomaly detection, evaluation beyond accuracy.

Only 0.17% of rows are fraud, which makes accuracy a meaningless metric. Demonstrate the key insight: "A model with 99.83% accuracy could simply predict not-fraud for every transaction and never catch a single case. Here is why accuracy is the wrong metric and what I used instead."

Fraud detection is a top use case in Pakistani fintech. Companies like NayaPay, SadaPay, and traditional banks are actively building these systems, so this project is directly hireable.

Advanced projects (month 4 to 6)

Project 9: End-to-end ML pipeline with deployment

Dataset: any of the above. Skills: MLflow experiment tracking, FastAPI model serving, Docker, model monitoring.

Take a completed model from projects 5 to 8 and productionise it: a FastAPI endpoint, tracked experiments, a Docker container, deployed to a free tier on Render or Railway. This is what separates data scientists who can build models from data scientists who can deploy them.

Document the full pipeline from raw data to live API, how you would monitor model drift, and what you would do if performance degraded after deployment.

Project 10: RAG system for document question answering

Dataset: any public document collection, National Assembly documents, State Bank of Pakistan annual reports, or PDFs in your domain. Skills: LangChain, vector databases, embeddings, RAG architecture, LLM integration.

Build a system that answers questions about a document collection with sourced answers. For example: "What does the State Bank of Pakistan say about freelancer foreign currency accounts?" answered by searching SBP circulars and citing the source.

RAG is the most in-demand applied AI skill in Pakistani software houses right now. A working RAG project on GitHub directly demonstrates what 10Pearls, Arbisoft, and VentureDive are hiring for. Stack: Python, LangChain, Chroma (free, local), and an LLM API free tier.

Project 11: Time series forecasting for business

Dataset: Store Sales dataset on Kaggle, or Pakistan monthly inflation data from PBS. Skills: time series decomposition, ARIMA, Prophet, time series feature engineering.

Forecast three months of retail sales, or Pakistan's monthly inflation. Time series needs different techniques from standard supervised learning, and proficiency here opens doors in finance, supply chain, and economic analysis. Document how you handled seasonality, whether your model captured trend changes, and what forecast uncertainty looks like.

Project 12: A full project with Pakistan-specific data

Dataset: Pakistan Labour Force Survey (PBS) or the Pakistan Social and Living Standards Measurement Survey (PSLM). Skills: complex cleaning, survey data analysis, weighted statistics, policy-relevant insights.

Answer a question that matters to Pakistani society. "What factors predict whether a Pakistani woman participates in the formal labour market?" Or "How does education level affect income across the four provinces?" Or "What is the relationship between household income and school enrollment?"

This is your strongest portfolio project. No candidate outside Pakistan will have it. It shows local domain knowledge, comfort with messy government survey data, and genuine curiosity about your country. Development economics is also a growing global field, so international organisations and research groups value this rigour too.

How to present your portfolio on GitHub

Your GitHub profile is your data science CV. Each project should follow a predictable structure:

project-name/
|- README.md          (the most important file)
|- data/
|  |- raw/            (original downloaded data)
|  |- processed/      (cleaned data)
|- notebooks/
|  |- 01_exploration.ipynb
|  |- 02_cleaning.ipynb
|  |- 03_modelling.ipynb
|- src/
|  |- utils.py        (reusable functions)
|- requirements.txt
|- results/figures/   (charts and visualisations)

Every README must have a one-paragraph summary of what the project does and why it matters, the question you answered, three to five key findings in plain English, embedded visualisations, simple instructions to reproduce your work, and a short note on what you would do differently with more time or data.

Also create a profile README (a file at github.com/yourusername/yourusername) that introduces you, lists your projects with one-line descriptions, and links to your LinkedIn.

Mistakes that make portfolios forgettable

Using Iris or Titanic as your primary project. Every student has analysed these. They signal that you completed a tutorial. Use them to learn, not to present.

Showing only accuracy. "93% accuracy" tells an experienced reviewer very little. Show precision, recall, F1, and a confusion matrix, and explain what each means in business context.

No business context. Every project should state what decision or action the analysis enables.

Dirty notebooks. Commented-out code, unrun cells, and no markdown is worse than no notebook. Run all cells top to bottom and explain each section.

Private repositories. If recruiters cannot see your code, your portfolio does not exist.

The realistic timeline

Month	Focus
Month 1	Projects 1 and 2. Learn pandas, matplotlib, basic cleaning.
Month 2	Projects 3 and 4. Strengthen visualisation and storytelling.
Month 3	Projects 5 and 6. Learn scikit-learn, evaluation, feature engineering.
Month 4	Project 7 or 8. Add NLP or advanced ML.
Month 5	Project 9 or 10. Add deployment or AI integration.
Month 6	Project 12 with Pakistan data. Polish all repos. Start applying.

Frequently asked questions

Do I need to buy a dataset to build a good portfolio? No. Every dataset mentioned here is completely free. Paid datasets are for production systems, not portfolio projects.

How many projects do I need before applying in Pakistan? Three solid, well-documented projects are enough to start. Three you can explain in depth beat ten you barely remember building.

Should I compete on Kaggle? Optional but valuable. Competitions give you interesting datasets, community notebooks to learn from, and a leaderboard position worth mentioning. Even a top 50% finish is worth noting on LinkedIn.

What programming language should I use? Python dominates Pakistan's data science job market in 2026. NumPy, pandas, scikit-learn, and Matplotlib are the foundations. Add PyTorch or TensorFlow for advanced projects.

Can I use AI tools to help build these projects? Yes, with the usual rule: use AI to write faster and understand better, not to build things you cannot explain. If Claude or ChatGPT writes a cleaning function, be able to explain every line in an interview.

Try a tool

Salary Calculator: see what a data science role pays after tax in Pakistan
Income Tax Calculator: calculate tax on any offer you receive

Browse all tools

All dataset sources listed here are verified as free and accessible as of July 2026. Availability and terms may change, always check the source platform before starting a project.