gemmaquiz

Type a subject, even misspelled, and get a quiz about it. The model runs on your machine through Ollama; the only thing that goes out is a Wikipedia lookup.

subject ──▶ gemma refines ──▶ you approve ──▶ wikipedia extract ──▶ gemma writes quiz ──▶ you play

How it works

1. A local Gemma model turns your rough input into a clean Wikipedia topic (revoltion french → French Revolution), and you approve it.
2. It reads the real Wikipedia article for that topic.
3. It writes a multiple-choice quiz from the article: 3, 5, 10, 15 or 20 questions.
4. You play, with instant feedback and a short explanation per question.

Grounding the questions in the article is the point: the model writes from real text instead of from memory.

Two settings worth knowing

• Default model gemma4:e2b. Benchmarked on an Intel Arc GPU it runs at about 28–32 tokens/s, versus about 16 for gemma4:latest: roughly 2× faster, and plenty good for quizzes. Set OLLAMA_MODEL=gemma4:latest for peak quality.
• Thinking is off. Gemma is a thinking model, but for structured JSON like topic cleanup and quiz generation, the reasoning adds latency without better results. OLLAMA_THINK=1 turns it back on.

Run it

ollama pull gemma4:e2b
git clone https://github.com/brbousnguar/gemmaquiz.git
cd gemmaquiz && npm install && npm start

Open http://localhost:3000. It binds to all interfaces, so other devices on your network can play too. The only runtime dependency is Express, and the frontend is vanilla JS with no build step.

Plain text: gemmaquiz.md · All projects