jevbooks

← All projects

litjev

zhengxuyu/litjev

Reproduction of Jev: turn any Qwen model into a fast decision model.

Scoring & Ranking26%Runner-up: Classification & RoutingMany questions, one request

What it is

LitJev is an independent research reproduction of the Jev decision layer on top of off-the-shelf Hugging Face Qwen checkpoints. Define questions and options, load a Qwen model, and get typed choices with probability distributions from one API call, with no training and no generated answer text.

How it uses Jev

LitJev serves the same POST /v1/systemone schema: a state plus questions of type choice, score and noul. Scores are read directly from the model's output head, so no answer text is generated. The answers object returns the winning option key, the probability-weighted level index, or the probability of yes, plus usage.

Primitives:choicescorenoul

Technique worth stealing

Read scores directly from the model's output head and serve typed questions over the public Jev schema, with no training and no generated text.

Try it

git clone the repo, cd litjev, then uv run --locked litjev --model Qwen/Qwen3.8-27B; open http://127.0.0.1:8000/.

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.96
  • Shows a System One patternno0.22
  • Handles uncertaintyno0.04
  • Measuredno0.06
  • Runnableyes0.78
  • Worth recommendingno0.30
  • Model replicayes0.96
  • Problem scopescore on a 0–2 scale0.87
  • About Jevyes0.98

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.