JEV-27B
AutoTrust's Apache-2.0 student of Jev 1.13 on Qwen3.8-27B, serving typed decisions and ordinary generation from one vLLM engine.
Overview
JEV-27B from AutoTrust is a LoRA distillation of TypeSafe Jev 1.13 onto Qwen3.8-27B, trained on the jev-distill-corpus-v3 dataset. It has a dual-head design: a System One path that answers noul, choice and score questions with calibrated probabilities, and a System Two path that leaves the base model's language head untouched for ordinary generation, both served from one vLLM engine. A smaller JEV-9B sibling exists.
The model card reports agreement with the Jev teacher on a 30k test split: mean KL 0.0186, noul AUROC 0.996, noul Brier 0.0013, score MAE 0.098 on a 0-5 scale, ECE 0.0009 after temperature, and choice top-1 agreement of 0.903 overall and 0.958 on decisive rows; the generation path scores 0.780 pass@1 on HumanEval. The AnotiaWang list adds that the authors report 84.07 versus 83.85 for Jev across six public decision benchmarks in their own runs. A live recommendation and search demo at news.guru.so ranks headlines and search results with calibrated decisions.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Yes/no question about a state | yes · no |
| Choice | Pick one of N named options | — |
| Score | Place the state on an ordered 0-5 rubric | — |
How it works
- Serve the model with vLLM
- Send Jev-format decision requests to the System One head
- Send ordinary prompts to the untouched generation path
- Use probabilities to rank, gate or route
Strengths
- Open weights reported to closely match Jev 1.13's distributions
- One engine serves both decisions and generation
- Apache-2.0 licence
Caveats
- 27B parameters need substantial GPU memory
- Benchmarks versus Jev are the authors' own runs
- Distilled from Jev, so it inherits Jev's known weaknesses
Getting started
Load autotrust/JEV-27B from Hugging Face with vLLM per the model card.At a glance
- Type
- Model
- Status
- beta
- Language
- Python
- Runtime
- vLLM / transformers
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
AutoTrust's Apache-2.0 model distilled from Jev 1.13 onto Qwen3.8-27B; one vLLM engine serves both the decisions and the untouched Qwen model for ordinary generation. The authors report 84.07 against 83.85 for Jev over six public decision benchmarks in their own runs. Smaller sibling: [JEV-9B](https://huggingface.co/autotrust/JEV-9B).
AnotiaWang/awesome-decision-models
Frequently asked questions
What is JEV-27B?
An open model distilled from Jev 1.13 onto Qwen3.8-27B that returns typed decisions with probabilities.
Can it also generate text?
Yes, the base model's generation path is preserved and served by the same engine.
How close is it to Jev?
The card reports mean KL 0.0186 against Jev targets and 0.903 choice top-1 agreement.
Related projects
Experimental 17M-400M English decision models for choice, yes/no, score and reranking, with weights, training code and an ONNX browser demo.
ModelresearchPythonMangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++