Module 14 explains how models are tested. Here you test them: the same 22 questions go to several small open models running on your device, and every answer is graded by code, not by eye. The questions check facts, honesty about what a passage does not say, following instructions, arithmetic, typos and attempts to hijack the model.
Contenders
Models
Run on
Leaderboardgraded by code, live on this device
Run the arena to fill this table.
Every questionclick a result to see the answer
Why grade with code
No Arguing with a Checker
Asking a person, or another model, whether an answer is “good” is slow and inconsistent. Questions with a checkable answer (a number, a format, a refusal) can be graded the same way every time, which is what makes scores comparable across models.
What a hallucination looks like
Confident, Fluent and Wrong
The passage never gives net income or a founding year. A model that answers anyway is filling the gap from patterns, not facts. Watch how often the smallest models do it, and whether bigger ones refuse.
Limits of this test
22 Questions Is a Small Exam
A model can pass a short test by luck and fail a long one. Real evaluations use thousands of items, several graders and repeated runs. Treat these scores as a demonstration of the method, not a ranking of the models.
Live, not recorded. SmolLM2 (Hugging Face), Qwen2.5 (Alibaba) and Llama 3.2 (Meta) are open models, downloaded from Hugging Face and run with ONNX Runtime in your browser. All questions and the passage were written for this page; Harbor Tools is fictional.