Typed decisions, in one page

A chatbot answer is a string you still have to interpret. A Laya answer is one of the options you already listed, plus a probability. That only helps when the decision was a classification to begin with: which queue, how urgent, whether someone asked for a refund.

ConvAI names the three shapes choice, score, and noul. Choice is a pick from your labels. Score is a position on a ladder you define. Noul is a single probability that a statement holds. The model card is the spec; this paragraph is only a map.

The number people quote from the specialist checkpoint, 0.766 on the authors’ 2,000-decision set, is not the zero-shot score of the base weights. The card says the base English and multilingual checkpoints land near a majority-class baseline on that suite. If you need those four workflows, load laya-typed-decisions or fine-tune. The upstream repo points at a Kaggle notebook for that job.

Two more limits are worth keeping next to any demo in the gallery. A long menu of labels shares one token budget, so a 70-way choice is a different problem from a 4-way choice. And a high probability is not calibrated until you fit a temperature on data that looks like yours.