mairamesniang12/aims-gaai-exam-wolof-assistant

# Real LLM Deployment and Standard Industrial Methodologies **Course:** Cases studies Applied Generative and Agentic AI **Assessment type:** Project-based exam **Theme:** From fine-tuning to real deployment **Recommended base model:** `Qwen/Qwen3-0.6B or litert-community/gemma-4-E2B-it-litert-lm.` This project is a continuation of the Session 7 fine-tuning lab. You will work in the same groups as for the case studies, as if you were working on a small industry-style AI project. The goal is not only to fine-tune a model, but to build a clean, reproducible, industry-style LLM deployment workflow. You must prepare data, train a LoRA adapter, evaluate the model, push the model to Hugging Face Hub, and deploy a working Hugging Face Space. ## Final Deliverables Each group must submit: 1. A GitHub repository containing the complete project. 2. A Hugging Face model repository containing the clean LoRA adapter. 3. A Hugging Face Space using Gradio or Streamlit. 4. A completed model card. 5. A short project report. 6. An individual technical note for each group member. 7. A short demo during the final presentation. ## Project Objective Build a small instruction-tuned LLM for a real use case under limited compute. Examples: - Wolof educational assistant; - health FAQ assistant with strict limitations; - agriculture advisory assistant; - public service assistant; - local language chatbot; - domain-specific assistant for AIMS coursework. The provided Wolof data is a starter example. Your group may keep Wolof, choose another African language, work in English or French, or build another domain-specific assistant. However, the methodology is mandatory: separated data sources, chat formatting, assistant-only training, evaluation, deployment, and documentation. ## Mandatory Methodology ### 1. Data Sources Must Be Separated Do not directly train on one mixed dataset without documenting the source. You must use at least three separated data sources: 1. a general or base …