Our summary
What if an ordinary language model chose an answer from a short list instead of writing a reply? Nish Tahir's tutorial explores that decision-model idea, also used by Jev. The model scores the allowed answers in one pass rather than producing words one at a time.
Tahir uses Qwen3-1.7B and assigns a letter to each possible answer. His program reads the scores for those letters and ignores the other words the model could produce. He then tests the answers, trains on examples and adjusts how sure the answers appear.
The untuned model looks too sure of itself: even a high score can belong to a wrong answer. The companion code covers testing, training and confidence adjustment. Tahir reports his own results, though, and this remains a tutorial rather than TypeSafe's model.
Key takeaways
- Use the scores for allowed answers without asking the model to write a response.
- Before trusting a score, check whether the answers that look certain are right.
- Train to change the answers; adjust confidence to change how sure they look. These are separate steps.