Live site · by pngwn

System One scorer

Qwen3.5-4B with a scoring head, type-safe by construction.

pngwn scores each state, question and option triple with a sequence-classification head on Qwen3.5-4B-Base and softmaxes the logits, so the output can only be one of the options supplied. The training code is in the repo. Held-out accuracy is 0.922 on AG News and 0.891 on Banking77.