Laya
Laya Overview
Laya is a multilingual, non-autoregressive System 1 decision model by Convai Innovations. Give it a state (text, email, ticket, or JSON document) plus typed questions (choice / score / noul) and it returns typed answers with mathematically calibrated probabilities in a single forward pass (~33 ms on T4) across 100+ languages. Trained with reinforcement learning against strictly proper scoring rules (RLCD), so reporting honest probabilities is the only way to maximize reward. It never generates text, so there is nothing to parse and nothing to hallucinate.
Three Checkpoints in One Repo (Apache 2.0)
| Checkpoint | Backbone | Params | Context | Best at |
|---|---|---|---|---|
| laya (repo root) | ModernBERT-large | 421M | 512 | English text, guardrails, email triage |
| laya-multilingual | mmBERT-base | 322M | 1024 (up to 8k) | 100+ languages, ~2.2x faster |
| laya-typed-decisions | ModernBERT-large | 421M | 1024 | four typed-decisions workflows (0.766 acc) |
Key Features
- Typed decisions: choice (up to 255 options), score (2-10 ordered levels), noul (yes/no), each with a full probability distribution.
- Speed: ~33 ms per question single (one forward pass, T4), 7.2 ms/question batched.
- RLCD training: reinforcement learning against strictly proper scoring rules; honest probabilities are what get rewarded.
- Jev-compatible: HTTP server via
pip install "laya[serve]"(laya-serve) and the same request format as TypeSafe Jev — drop-in migration path. - Router: picks the right checkpoint per request and keeps two checkpoints resident.
- Fine-tuning: Kaggle free 2xT4 notebook covers dataset build, RLCD training, temperature fitting, evaluation, and Hub push.
Laya vs Jev (TypeSafe AI)
| Dimension | Laya | Jev |
|---|---|---|
| Availability | Open weights, Apache 2.0, self-host | Closed commercial API |
| Params / backbone | 421M ModernBERT-large / 322M mmBERT-base | Undisclosed |
| Languages | 100+ (multilingual checkpoint) | Not specified publicly |
| Latency / cost | ~33 ms, free when self-hosted | 0.114 s / $0.000081 per decision; $42 per B input tokens |
| typed-decisions | 0.766 fine-tuned (base zero-shot ~0.36) | 0.727 published |
| Banking77 (77 options) | 0.425 at default settings (raise head_max_len) | 0.870 |
| Calibration | Ships over-confident (ECE 0.466); refit temperature → 0.081 | Calibrated confidence marketed natively |
Honest Limits
- Base checkpoints are near chance zero-shot on typed-decisions (0.362 vs 0.318 random, 0.461 majority) — fine-tune on your domain before relying on it. Laya is a fast base to specialize, not a zero-shot decision engine.
- noul answers can follow the option labels instead of the state (issue #156);
act_probabilitycarries no usable signal yet (issue #185) — gate onconfidenceinstead (AUROC 0.77). - 77+-option choice questions need a bigger head budget (head_max_len=512, max_len=1024) or accuracy falls off sharply.
- Root checkpoint is English only; use laya-multilingual for anything else.
Installation
pip install laya · pip install "laya[serve]" for the Jev-compatible HTTP server · demo: https://huggingface.co/spaces/convaiinnovations/laya-demo · GitHub: https://github.com/NandhaKishorM/laya