Laya does not take a prompt in the conversational sense. It takes a structured question and answers it once. Everything you build sits on three primitives.
Choice — classify
You supply a criteria table of options; Laya returns probabilities per option plus a confidence value.
Good fits:
- Ticket routing to one of five teams
- Intent labelling
- Sentiment buckets
Keep the option list short. Accuracy degrades as the number of options grows — this is the model's clearest weakness, so prefer staged decisions: first route to a department, then classify within that department.
Score — rate
Where the answer is a number rather than a label: risk level, priority, quality band.
Two habits matter here:
- Pin the scale. Make the rubric explicit in your training data, or the model learns your labelers' inconsistencies.
- Never automate on the raw score. Fit temperature first, then set a threshold that is informed by the calibrated probabilities.
Noul — boolean
A single yes/no decision, designed for guardrail work: prompt-injection detection, policy checks, spam gating.
The structural advantage is worth stating plainly. Because the model never generates text, there is no output channel through which injected instructions can be executed. A generative model asked to classify a malicious email may read the email's instructions as its own. Laya cannot.
Modelling advice that saves weeks
One primitive per question
If you find yourself asking "which category is this, and how risky is it, and should we block it", that is three calls. Splitting them gives you independent calibration and independent thresholds.
Batch related judgements
Multiple questions can be resolved in a single request over the same input, which is where per-decision cost drops furthest.
Watch the language router
The English checkpoint behaves badly on non-Latin scripts — and it fails *confidently*. Always let the router assign the language before inference.
A worked example
Consider an inbound contact form. Three primitives, one input:
1. choice → route to sales / support / abuse 2. score → priority 0–100 3. noul → does this need immediate human review
Each returns its own calibrated probability. You set three separate thresholds, and you can tune any one of them without retraining the others.
Design the decision tree before you touch the model. Most disappointing results come from asking one question that should have been three.