A small Python example scores choices without generating an answer

NobodyWho demonstrates a local classifier that reads a language model’s next-token scores for a few allowed labels, then converts those scores into a distribution. The Python example uses a small Qwen model rather than generating a prose answer and parsing it. Its Jev-themed presentation is explicitly a parody, not an implementation of TypeSafe’s model or training method.

NobodyWho · Sep 22, 2026

Why it matters

The example exposes a useful mechanism while making the difference between a choice score and reliable confidence worth examining.

More context

Try the educational example with Python 3.12+, llama-cpp-python and the downloaded GGUF weights. Evaluate it on labelled examples before using it for decisions; local execution still needs compatible hardware.

Keep in mind

It provides no equivalent calibration training or evidence of matching Jev’s quality. Normalizing a few token scores does not make them calibrated probabilities of being correct.

Original read