LLMs for Arabic and Darija

Make language models work for Arabic and Algerian Darija: measure them, fix their weak spots, and build an app your users understand.

Most language models are built for English. You learn why they struggle with Arabic script, Darija, and the French-Arabic mix people really write, then build a benchmark to measure it. You finish with a working assistant tuned for Darija, using prompting, retrieval, and light fine-tuning.

What you will be able to do

• Explain why LLMs handle Arabic and Darija worse than English
• Build a small benchmark to compare models on Darija
• Ship an assistant that understands Darija, Arabic, and French

Before you start

• Python basics
• You read Arabic or Darija

Course outline

1. Why models struggle: A tokenizer comparison on Arabic, Darija, and French. A cost and quality note per language.
2. A Darija benchmark: A small test set with code-switching. A model comparison you can share.
3. Prompting and RAG for Arabic: An assistant grounded in Arabic documents. Before and after scores.
4. Light fine-tuning: A small model tuned on Darija examples. An assistant that understands Darija, Arabic, and French.

All AI and machine learning courses