What it does
The trained head beats the same backbone's generation baseline 95.8% to 80.0% on the bundled suite, and everything runs on a laptop CPU. Weights and data are public.
Benchmarks & research
minojev
Freeze a small Qwen model, train a decision head smaller than a million parameters, and it answers typed questions with zero generated tokens.
Bookmark: minojev
Keep this in your collection.
Leave a noteWhat would you try with this?
: minojev
Only you can see your notes.
Screenshot unavailable. Open the experiment ↗
How you can use it
You could build a puzzle game where a character decides its next safe move. Instead of waiting for an AI to write words, your app provides a list of moves. The software immediately picks the best option without typing a response.
A developer can install the software to run offline on a regular computer. They can test the included maze project to see how the model chooses a path. Your app must provide at least two choices, and no more than 255, for every question.