Laya

Laya is an open-source family of decision models released by Convai Innovations under the Apache-2.0 licence. Given a state (a message, email, ticket or JSON object) and a set of typed questions, it returns typed answers with a probability for every option, in a single forward pass and without generating text.

How it works

• Backbone: a bidirectional encoder. The English checkpoint uses ModernBERT-large (421M parameters in total, 512-token context); the multilingual one uses mmBERT-base (322M, 1,024-token context, 100+ languages).
• Decision head: two transformer layers, an option-marker scorer that scores each option at its own [MASK] token, and an act/escalate head.
• Question types: choice, score and noul. Options are defined at request time, so new schemas need no retraining.
• Training: RLCD, reinforcement learning against strictly proper scoring rules, so honest probabilities maximise reward.

A third checkpoint, laya-typed-decisions, is ModernBERT-large fine-tuned on four specific workflows; see typed decisions.

Reported performance

On a Tesla T4, one question takes 39.5 ms (English) or 32.8 ms (multilingual). Headline accuracy figures from the model card include AG News 0.950 and, for the fine-tuned checkpoint, typed-decisions 0.766.

Honest limits (from the model card)

• Base checkpoints are near chance on the typed-decisions benchmark zero-shot; the strong result comes from fine-tuning.
• Large label sets (50+ options) degrade at default token budgets; Banking77 scores 0.425.
• score is the weakest question type, and noul can follow its false/true labels instead of the state.
• The checkpoints ship over-confident; fit temperature scaling on your data.
• action.act_probability carries no usable signal yet.

Laya Studio hosts Laya behind the Jev-compatible /v1/systemone API and is independent of Convai Innovations. Start with what is Laya.