mohamedaminehnioua/MedAtlas

A Moroccan-Darija medical assistant A Moroccan-Darija medical assistant — fine-tuned on ~92k cleaned Darija medical Q&A pairs, served as a chat API behind a simple web app. --- Med·Atlas takes a raw, messy dataset of Moroccan-Darija medical questions and answers, cleans it, fine-tunes Atlas-Chat-2B (a Gemma-2 based Darija model) on it, and exposes the result as a chat assistant that answers health questions in Darija. The whole flow lives in four notebooks — **explore → clean → train → serve** — plus a static web front-end. > ⚠️ Med·Atlas is an educational project. It does **not** replace a doctor. ## The pipeline at a glance | Step | Notebook | Output | |------|----------|--------| | 1. Explore | `med_qa_ma_eda.ipynb` | Data-quality scorecard | | 2. Clean | `clean_dataset.ipynb` | `med_qa_ma_clean.csv` (92,282 rows) | | 3. Train | `train_atlas_chat_2b.ipynb` | LoRA adapter for Atlas-Chat-2B | | 4. Serve | `med_atlas_colab.ipynb` | Public chat API (`/chat`) | | 5. Use | `web/` | Darija chat web app | ## Project structure ``` ├── data/ │ ├── med_qa_ma.csv # raw dataset (~101k Q&A pairs) │ └── processed/ │ ├── med_qa_ma_clean.csv # cleaned dataset (92,282 rows) │ └── removed_rows.csv # every dropped row + the reason (audit trail) ├── notebooks/ │ ├── med_qa_ma_eda.ipynb # 1. explore & judge the raw data │ ├── clean_dataset.ipynb # 2. reproducible cleaning pipeline │ ├── train_atlas_chat_2b.ipynb # 3. QLoRA fine-tune Atlas-Chat-2B │ └── med_atlas_colab.ipynb # 4. serve the model as a chat API ├── web/ # static chat front-end (HTML / CSS / JS) │ ├── index.html │ ├── app.js │ ├── styles.css │ └── assets/ # logos └── docs/assets/ # charts shown in this README ``` ## 1 · The data The raw dataset (`med_qa_ma.csv`) has **100,966 rows** and three columns: - **Question** — patient question, in Moroccan Darija - **Answer** — reply, in D …