Why this matters
Most text-processing systems either generate free text (risking parsing and hallucination) or treat classification as independent monolithic tasks. Laya Multilingual swaps generation for typed, structured decisions: give it a state (email, ticket, JSON) and a schema of typed questions and it returns typed answers plus calibrated probability distributions in one forward pass. That shift lets teams build deterministic branching logic on honest confidences while keeping latency low across 100+ languages.
Key Capabilities
- Single-pass typed decisions: answers for multiple questions (choice, ordinal score, noul boolean probability) are produced in one non-autoregressive forward pass, avoiding text generation and parsing errors. This is useful for routing, triage, moderation, and other high-throughput classification tasks.
- Multilingual and long-context support: built on an mmBERT-base backbone and a small decision head to support 100+ languages and up to 8,192 tokens (configurable). Benchmarks show substantially better cross-lingual robustness than an English-only checkpoint and faster per-question throughput at scale.
- Calibrated decision training (RLCD): trained with reinforcement‑learning objectives that encourage honest probability estimates rather than overconfident point predictions. The architecture scores options at dedicated markers so answer spaces are defined per request with no retraining.
Who it's for, and tradeoffs
Great fit if you need low-latency, high-volume multilingual routing/triage/classification where structured, machine-readable answers and honest confidence scores let you build deterministic downstream logic. It’s also useful when you want a smaller, GPU-efficient decision model (≈322M params) rather than a generative LLM.
Look elsewhere or validate carefully if your workload relies heavily on English-only fine-grained ordinal scoring, extreme label cardinalities (>~20 choice options), low-resource languages with sparse training data, or out-of-distribution long-document accuracy without in-domain tuning. Practical cautions: the checkpoint ships uncalibrated by default (refit per-question-type/option-count temperatures on held-out data), ordinal score outputs show a measured position bias, and noul probabilities can under-report true positives—use choice fallbacks or local recalibration where necessary.
Where it fits
Position this as a production-grade System‑1 decision layer that complements generative LLMs: use Laya for deterministic routing, guardrails, and classification with honest confidences, and reserve generative models for open-ended text synthesis or complex reasoning. For English-dominant workloads, pair with the English Laya checkpoint and use the Router to dispatch by script/language.
