Convai's open decision family: one forward pass, Choice / Score / Noul, Apache 2.0. Not a tiny chatbot wearing a trench coat.
Publisher and license
Laya is released by Convai Innovations under Apache 2.0 on Hugging Face (convaiinnovations/laya). The upstream SDK and training code live at github.com/NandhaKishorM/laya with a PyPI package named laya. This directory is independent. We do not represent Convai.
Checkpoints in the family
Three public checkpoints share one hub repo: the English root model (ModernBERT-large, about 421M params, 512 context) for English guardrails and email-style triage; laya-multilingual (mmBERT-base, about 322M, 1024 context, up to 8k encoder RoPE) for 100+ languages; and laya-typed-decisions (ModernBERT-large, 1024 context) fine-tuned for four workflow-style task groups Convai documents on the card.
Only the subfolder you request downloads.
Primitives: choice, score, noul
Questions are declared as JSON. Choice picks among criteria keys with softmax over option markers. Score returns an ordinal rating. Noul returns a calibrated probability for a yes/no style statement.
This mirrors the TypeSafe Choice, Score, and Noul vocabulary in /learn/jev-typesafe. Useful for mental mapping, not a promise of API compatibility.
Router mode
Router. It detects script and language in sub-millisecond Python and routes to English or multilingual checkpoints. Convai documents large accuracy gaps when English-only weights see non-Latin scripts, so routing is a safety feature, not just convenience.
Preload checkpoints in memory to avoid multi-second cold rebuilds when languages alternate.
Training objective
Convai describes RLCD (reinforcement learning for calibrated decisions). The policy is rewarded with strictly proper scoring rules so honest probabilities maximize expected reward. That is why teams want System One models instead of parsing chat text.
You still need domain temperature fitting on your data before you trust thresholds.
Honest limits (from the model card)
Base English and multilingual checkpoints are weak on Convai's typed-decisions benchmark without fine-tuning. 766 accuracy figure belongs to laya-typed-decisions, not the root weights.
High-cardinality choice questions can exhaust per-option token budgets unless you raise head_max_len or split hierarchically. Ordinal score tasks are weaker than choice/noul on some public sets.
Do not run English-only root for non-English production without multilingual or routed paths.
More in this hub
Related Jev learn guides
Live demo
Lonely__MH's laya-mlx snake clip shows local typed decisions driving a game loop on Apple Silicon. Not a benchmark, but a useful feel for throughput.
FAQ
- Is Laya generative AI?
- No. It does not continue text token by token. Outputs are discrete decisions and probabilities over your schema.
- How do I install it?
- pip install laya, then laya.load() or Router(preload=True) per the Hugging Face quickstart. Set USE_TF=0 if Transformers TensorFlow probing hangs your environment.
- Is Laya the same as Jev?
- Same product category (typed parallel decisions), different vendor, license, and hosting model. Compare in /learn/laya-vs-jev/compare.