Data Science Roadmap

A 31-module open-source data-science course — from Python basics through transformers, diffusion, and production LLM workflows. Free, MIT-licensed, every notebook one click away in Google Colab.

What it is

An open-source self-study course that walks from print("hello") through working transformers, diffusion models, RAG pipelines, and production LLM workflows. 31 deep-dive Jupyter notebooks, every one runnable in Google Colab in one click, organised into six parts so concepts compound module to module.

The course is built around the principle that learning lands when you can read it, run it, edit it, and re-run it. Every concept arrives as runnable code first, with a short companion explanation of why it works that way — no videos to scrub through, no graded quizzes, no paywalls.

Why I built it

Most “learn data science” tracks fail in one of two predictable ways:

• They split material across three or four separate paid courses (statistics, then Python, then ML, then “AI”) that don’t talk to each other and leave you patching gaps for months.
• Or they jump from print("hello world") straight to a Kaggle notebook with no in-between, leaving anyone who’s not already comfortable in Python stranded.

This repo is the in-between — one continuous path, free, with the modern AI track that working ML engineers actually need (RAG, fine-tuning, MLOps, evals) included rather than punted to a separate “advanced” course.

Course structure

Part	Track	Modules
1	Python for Data Science	01 → 05
2	Data Visualization	06 → 10
3	Data Analysis & ML Foundations	11 → 16
4	Machine Learning & AI (deeper dive)	17 → 22
5	AI-Research Foundations	23 → 25
6	Practitioner Skills	26 → 31

Part 1 — Python for Data Science (M01-05)

Variables, data structures, OOP, file I/O, NumPy, Pandas, APIs, web scraping. The alphabet of every later module.

Part 2 — Data Visualization (M06-10)

Matplotlib’s object-oriented API, the seven core chart types, specialised viz (waffle, word cloud, Folium maps), animation and Plotly, and dashboards that tell one cohesive story.

Part 3 — Data Analysis & ML Foundations (M11-16)

The universal data-science workflow — import → wrangle → explore → model → evaluate → communicate — built around a shared dataset (auto-mpg) so each module compounds on the last. The capstone (M16) deliberately switches to California Housing to prove the workflow generalises end-to-end.

Part 4 — Machine Learning & AI (M17-22)

PyTorch fundamentals; the six core model archetypes (Linear, Logistic, K-Means, MLP, CNN, Transformer LM); self-attention implemented from scratch with d_model = 2 so every matrix is hand-checkable; multi-head + causal attention; diffusion models on a 2D toy; time-series forecasting with ARIMA, Prophet, and LSTM.

Part 5 — AI-Research Foundations (M23-25)

The math under every neural network (functions, derivatives, gradients, matrices, probability) integrated with the code, a deep PyTorch primer, a guided dissection of DeepSeek-V3’s actual inference code (RMSNorm, RoPE, Multi-Latent Attention, Mixture-of-Experts), and worked fine-tuning examples — full fine-tuning, LoRA, QLoRA, and SFT with TRL.

Part 6 — Practitioner Skills (M26-31)

The day-to-day skills working data scientists and ML engineers use but most courses skip:

• SQL — JOINs, CTEs, window functions, the SQL ↔︎ Pandas bridge
• Tree-based models — Random Forest, XGBoost, LightGBM, SHAP for interpretation
• A/B testing — proportion z-test, sample-size calc, Bonferroni / BH correction, the peeking trap
• MLOps — FastAPI, Docker, MLflow, drift monitoring with KS + PSI
• RAG & vector search — embeddings, Chroma, hybrid BM25 + vector, reranker, grounded answers
• Prompt engineering & LLM eval — few-shot, chain-of-thought, ReAct, structured outputs, LLM-as-judge

What makes it different

	This course	Typical course
Production architecture depth	DeepSeek-V3 inference dissection	“Transformers exist”
Math integrated with code	Yes, in Module 23	usually skipped or paywalled
Practical skills (SQL, A/B, MLOps)	Modules 26–29	rarely covered
Companion docs	Line-by-line, colour-coded, ~30 pages each	None
Cost	Free, MIT-licensed	$40–300/month

Audience

• Beginners who want a structured path from “what is a variable” to a working ML model — without paying for a Coursera certificate.
• Self-taught coders filling gaps in Pandas, NumPy, and scikit-learn — and reaching beyond into PyTorch, transformers, and RAG.
• Career-switchers building a portfolio. The capstone (Module 16) and the production modules (M29-31) are portfolio-ready as-is.

What you can do at the end

After the 31 modules you can:

• Write any Python program and load data from any source (CSV / Excel / JSON / SQL / API / scraped HTML).
• Clean, explore, visualise, and communicate findings on real datasets.
• Build classical ML models (regression, gradient boosting) AND modern AI models (transformers, diffusion).
• Read production LLM source code (DeepSeek, Llama, Mistral, Qwen) and recognise every architectural piece.
• Fine-tune any open-weight model on your own data with LoRA / QLoRA.
• Ship a model behind FastAPI + Docker with MLflow tracking.
• Build a working RAG pipeline with vector search and grounded answers.
• A/B-test prompts and evaluate LLMs scientifically.

That’s effectively a 2026 ML-engineer career, built up step by step from print('hello').

How to use it

Option A — Colab (zero install). Click any Colab badge in the README, hit File → Save a copy in Drive, run the cells.

Option B — Locally.

git clone https://github.com/kader-xai/data-science-roadmap.git
cd data-science-roadmap
pip install jupyter numpy pandas scikit-learn torch transformers
jupyter notebook

Tech stack

• Language — Python 3.10+
• Core libraries — NumPy, Pandas, scikit-learn, Matplotlib, Plotly
• Deep learning — PyTorch, Hugging Face Transformers, PEFT, TRL
• Production track — FastAPI, Docker, MLflow
• Retrieval & LLM track — Chroma, BM25, BGE embeddings, Ollama
• Statistics & boosting — SciPy, XGBoost, LightGBM, SHAP
• Format — Jupyter notebooks (one per module) + companion Markdown explanation docs

Links

• 💻 Repo: github.com/kader-xai/data-science-roadmap
• 🌐 Live course site: kader-xai.github.io/data-science-roadmap
• ▶️ Start with Module 1 in Colab: open notebook
• 📝 Blog write-up: Data Science Roadmap — From print(‘hello’) to Production LLMs
• 📦 Related project: Employee Recall — LoRA + RAG
• 📜 License: MIT