Laya

An encoder.config.json

A BERT-family transformer that reads the whole input in both directions at once and produces no text at all. config.json says ModernBertForMaskedLM: hidden size 1024, 28 layers alternating full and sliding attention, mean pooling.config.json

That is the sharpest contrast with a chat model. GPT-style models are decoders: they predict the next token, then the next, building a sentence. An encoder does not predict the next anything. It reads, and it scores. One forward pass, about 33 ms, no tokens produced.model card

Full build: ModernBERT-large as the backbone, fully fine-tuned, plus a decision head trained from scratch (two transformer layers, an option-marker scorer, and an act/escalate head). 421M parameters in total.model card

512 tokens per question on the English checkpoint, 1,024 on the multilingual one.model card

Separately, the decision head has its own budget for the options: head_max_len is 192 tokens on English, 256 on multilingual. That number matters, and question 6 explains why.model card

Every question in a call is answered in a single forward pass.model card

Every option you send gets its own [MASK] token.model card

The model scores each of those markers, then softmaxes across them, and that distribution is the answer. This is the part worth reading twice.model card

Which means the answer space is defined at request time, not at training time. The options live in your request, so new schemas need no retraining.model card

That is how a model trained on support tickets can answer LANE_LEFT versus IDLE on a highway: nobody taught it to drive, they taught it to read a situation and score the markers you supply.

RLCD, Reinforcement Learning for Calibrated Decisions, with the method published.model card

The policy reports a probability distribution. Exploration adds zero-mean Gaussian noise to the logits. The reward is a strictly proper scoring rule (log and spherical, plus ranked probability score for ordinal questions). Updates are REINFORCE with a group-mean baseline, the same idea as GRPO.model card

The phrase doing the work is strictly proper. Under such a rule, the only way to maximise expected reward is to report the probabilities you actually believe. Overclaiming is punished, so honesty is the optimal policy rather than a good intention.

TypeSafe uses the same name for Jev's training without publishing any of this.our inference, from the absence of a published method

Price

Free. Apache 2.0. You pay for hardware, and nothing leaves your machine.model card

Install

pip install laya. Weights, about 2.3 GB, download from Hugging Face on first use.

Speed, stated

39.5 ms for one question on a T4 GPU, 32.8 ms on the multilingual checkpointmodel card

Speed, measured

p50 107 to 185 ms across the three games, on a laptop CPU with no GPUmeasured here

• Handle many options well.The options share that fixed head_max_len budget, so 77 options get roughly 3 tokens each and the labels blur together. On Banking77, Convai report Laya at 0.425 against Jev's 0.870.model card The limit is a consequence of the mechanism in question 3, not a mystery.
• Match Jev on soft distributions.Higher argmax accuracy on typed decisions (0.766 against 0.727) but lower soft accuracy (0.471 against 0.580).model card
• Arrive calibrated out of the box.Its 0.081 ECE comes after domain temperature fitting. Raw, the base checkpoint is worse than Jev.model card