18 notebooks that actually teach data science by doing

What it does

This repo is a structured curriculum of 18 standalone Jupyter notebooks, organized into four difficulty levels. Each project pairs a real dataset with a specific technique — pandas and GridSearchCV at the bottom, YOLOv8 and Detectron2 at the top. The goal is to replace tutorial videos with runnable code you can break and fix.

The interesting bit

The sequencing matters. Level 1 is deliberately unglamorous — Titanic, Iris, customer churn — because the author treats “getting comfortable with messy data” as a skill worth isolating. By Level 4 you’re doing network intrusion detection and industrial robotics imaging, but the scaffolding is meant to make that jump feel incremental rather than performative.

Key highlights

• 18 projects across classification, regression, NLP, computer vision, anomaly detection, and robotics
• Each project ships with its own requirements.txt and is designed to run after cloning
• Explicitly covers both “from scratch” implementations (logistic regression, sentiment analysis) and modern tooling (EfficientNetV2, ResNet34, LLMs)
• Includes medical imaging (eye disease, Alzheimer detection) and safety-critical applications (cyber-bullying prediction)
• MIT licensed; accepts contributions with a one-project-per-PR rule

Caveats

• The “Advanced” level is only 4 projects deep — ambitious readers will outgrow it quickly
• Large datasets must be hosted externally; some projects may require manual dataset hunting
• No visible CI or automated testing to verify notebooks still execute cleanly

Verdict

Solid for students, career switchers, or anyone who learns by breaking code. Skip it if you already have a portfolio — this is a starting line, not a differentiator.