Skip to content
JevDirectory.orgTraffic count unavailable. View site traffic
API & Models

Decision head

The small trainable module added to a frozen backbone that turns hidden states into one probability per option.

All 164 terms
Category
API & Models
Also known as
—
Related terms
4
Directory entries
2
Docs
—
Added
2026-10-10

Definition

AutoTrust's Blocks-of-Experts recipe pairs a frozen Qwen backbone with a rank-16 LoRA plus a 24-slot decision head; only 108.9M of JEV-27B's 27B parameters train, about 0.4% of the model. The head is initialized from the backbone's own lm_head rows, so at step zero it already matches the pretrained restricted distribution.

A head is cheap to train and cheap to carry. JEV-9B's 40.2M-parameter block took roughly 3 B200-hours and left HumanEval pass@1 at 70.7%, where folding the adapter into the backbone would have dropped it to 61.6% — which is why the block ships separate.

Tagsmodelsarchitecture

Related terms

Definitions that connect to this one.

Option markerScoring each answer option at its own marker token, then softmaxing over the question's options so the answer space is defined at request time.API & ModelsEncoder backboneThe bidirectional transformer under a decision model that reads the state and options once, rather than generating tokens.API & ModelsDistillationTraining a smaller model to reproduce a larger model's output distribution, including its probabilities rather than labels alone.Evaluation & TrainingLogit readoutReading the next-token probability of each declared label, then renormalising, so no answer text is generated.Patterns

From the Jev directory

AutoTrust JEV-27B
An open distillation of Jev 1.13 onto Qwen3.8-27B: a 108.9M-parameter decision block answers typed questions in one prefill pass and adds a POST /v1/decide route beside a full System 2.
Jev Models & Reimplementations#community#open-models#local
Community
AutoTrust JEV-9B
The 9B member of AutoTrust's Jev students: a 40.2M-parameter decision block on Qwen3.5-9B with 0.019 mean KL to Jev 1.13 and 70.7% HumanEval left untouched, small enough for one modest GPU.
Jev Models & Reimplementations#community#open-models#local
Community

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai