Laya

Open-source (Apache 2.0, claimed in the README) decision model by the GitHub user NandhaKishorM (repo NandhaKishorM/laya, weights under convaiinnovations on…

Laya

Open-source (Apache 2.0, claimed in the README) decision model by the GitHub user NandhaKishorM (repo NandhaKishorM/laya, weights under convaiinnovations on Hugging Face). It answers typed questions (choice, score, noul = P(true)) about a piece of text in one encoder forward pass, with calibrated probabilities and no text generation. It is positioned as the open-weights, self-hostable counterpart to TypeSafe Jev, a proprietary hosted decision API; the server speaks Jev’s wire protocol (POST /v1/systemone), so Jev clients can switch by changing a base URL. “Jev aka Laya” is therefore a compatibility relationship, not one project under two names.

Facts

Kind	Encoder (ModernBERT-large 421M English; mmBERT-base 322M multilingual) with a trained decision head. Not an LLM.
Install	pip install laya (Python 3.10+); extras for serve, mcp, langchain, llamaindex, crewai, onnx, fast, structured. TypeScript: laya-ts.
Core API	Router().predict(state, questions); the router picks a checkpoint by script and language.
Checkpoints	laya (English), laya-multilingual (100+ languages, up to 8,192 tokens with max_len=8192), laya-typed-decisions (fine-tuned).
Training idea	RLCD: reinforcement learning against strictly proper scoring rules so honest probabilities maximise reward.
Speed (author, T4)	about 33 ms for one question; 103-332 questions/s batched.

Detail, benchmark provenance and the full list of limits: Laya.

When it fits (and when not)

• Fits: high-volume, latency-sensitive routing or triage where the label set is small (under about 20 options), you can fine-tune on your own labelled decisions, and you want a probability you can threshold, without an LLM call.
• Does not fit zero-shot: the base checkpoints score below the majority-class baseline on the typed-decisions benchmark. The headline accuracy is for a checkpoint fine-tuned on that benchmark’s training split.
• Does not fit: 50+ option choice questions without tuning (Jev wins there by the author’s own numbers), negated or contrastive text, or use as a security control for prompt injection (0.698 held-out, n=116).

Operational notes

• laya-serve binds 0.0.0.0 with no authentication unless LAYA_API_KEY is set.
• laya-multilingual ships with no fitted temperatures; fit them before trusting its probabilities.
• Ported Jev thresholds do not transfer: confidence is defined differently.

Related

• Laya — full claim-by-claim summary with provenance flags
• Model Tier Routing — routing work to LLM tiers (different thing; Laya’s model-router preset can feed such a decision)
• LLM-as-Judge — a generating judge, the opposite design point