Large Language Models in Data Science (WIP)

This repository contains materials for a 21-hour lecture titled:

Large Language Models in Data Science
Embeddings, Prompting, Retrieval & Applied Use Cases

---

🎯 Course Objectives

• Understand how modern LLMs work: tokens, embeddings, transformer architecture, and inference.
• Use pretrained models via Hugging Face and API-based LLMs in a reproducible way.
• Design effective prompts and retrieval-augmented generation (RAG) pipelines that use embeddings and keyword search.
• Compare different approaches (regex rules, classical ML, HF models, zero-shot LLMs, RAG) in terms of quality and latency.
• Evaluate model performance with simple but meaningful metrics (accuracy, latency, precision/recall, grounding) and apply LLMs to real data science tasks (analysis, reporting, feature generation, code assistance).

---

📚 Structure

• Language: English
• Audience: Final-year data science students with strong math background
• Total Time: 21h
  • Lectures: ~7h
  • Labs: ~7h
  • Hackathon: ~7h (final evaluation)

---

💻 Content

Folder	Description
data/	Data sets
labs/	Jupyter notebooks for hands-on experimentation
slides/	LaTeX slide decks for all lectures

---

🧪 Final Evaluation: Hackathon

The final mark (out of 20) will be based on participation in a group hackathon (teams of 1 to 5 participants). Each team will present a project that uses LLMs in a data science context.

🔍 Evaluation Breakdown (4 x 5 points)

Criterion	Description
1. Use Case & Applications (5 pts)	Relevance, originality, and potential future usefulness of the proposed use case
2. Code Cleanliness (5 pts)	Code readability, organization, comments, modularity, and reproducibility
3. Lecture Integration (5 pts)	Use of concepts covered in the course: prompting, fine-tuning, RAG, etc.
4. Presentation (5 pts)	Clarity, conciseness, visuals, and ability to explain design decisions

💡 Your final score out of 20 is your official course grade.

---

📅 Timeline

Materials for Weeks 1–6 (lectures, labs, and hackathon template) are available in this repository. Future iterations may refine content and add additional examples.

---

🗓️ Weekly Materials

Week 1 – LLMs: A Primer

• Slides: slides/week_1_intro.tex
• Lab: labs/week_1_tokenization.ipynb
• Readings: The Illustrated Transformer, OpenAI GPT Models Documentation, Attention Is All You Need

Week 2 – Hugging Face & Transformers: Using Pretrained Models

• Slides: slides/week_2_huggingface.tex
• Lab: labs/week_2_huggingface_basics.ipynb
• Docs: Transformers Overview, Datasets, Hugging Face Hub

Week 3 – Effective LLM Use: Coding, Research, Ideation, First Analysis

• Slides: slides/week_3_effective_llm_use.tex
• Lab: labs/week_3_effective_llm_use.ipynb
• Notes: The lab uses the Gemini API. Set an environment variable GEMINI_API_KEY (e.g., add it to .env) and install google-generativeai before running.

Week 4 – Text Classification and Intent Routing

• Slides: slides/week_4_classification.tex
• Lab: labs/week_4_classification.ipynb
• Notes: Build and compare several intent classifiers (regex rules, embeddings + logistic regression, zero-shot NLI, LLM prompting) for an AMU student chatbot; analyse accuracy and latency trade-offs.

Week 5 – Retrieval-Augmented Generation (RAG)

• Slides: slides/week_5_rag.tex
• Lab: labs/week_5_rag.ipynb
• Notes: Construct a small RAG system over an AMU / Data Science corpus; explore BM25 vs. embeddings vs. hybrid retrieval; observe how augmented prompts change answers and reduce hallucinations.

Week 6 – Hackathon: Applied LLM Projects

• Slides: slides/week_6_hackathon.tex
• Template: labs/week_6_hackathon_template.ipynb
• Notes: One-day team hackathon (1–5 students) to design and implement a small end-to-end LLM/RAG project in a GitHub repository, then present a short demo. See the “Final Evaluation: Hackathon” section above for grading criteria.

---

📬 Contact

Instructor: Sebastian Mueller
University: Aix-Marseille Université
Course code: TBD