What it does
The repository contains a preregistered experiment, results and a teaching interface for CUDA or Apple Silicon.
Benchmarks & research
mini-jev
Mini-Jev reads option-letter logits from a frozen Qwen3-4B instead of asking the model to generate JSON.
Bookmark: mini-jev
Keep this in your collection.
Leave a noteWhat would you try with this?
: mini-jev
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
How you can use it
Your developer can use this method to help your app sort information faster. You give the AI a text and a short list of choices. Normally, an AI types its answer word by word. This tool simply checks which single option letter the AI already prefers.
Your developer can run the included teaching page to test this on your own questions. The fast method only works for lists with up to twenty-six choices. If your app needs the AI to write new sentences, it must still type them out. Testing requires an Apple or NVIDIA computer.