jevbooks

← All projects

diffusion-jev-sglang

Hangzhi/diffusion-jev-sglang · Homepage

Visual decisions with DiffusionGemma and SGLang.

Classification & Routing75%Runner-up: Scoring & Ranking

What it is

A visual classification demo that selects doodle and flower labels from image pixels using typed Choice questions, with a drawing playground and public evaluation artifacts. For developers exploring Jev-like decision engines.

How it uses Jev

Jev-style adapter maps options to letters, reads answer-letter scores from DiffusionGemma's last active denoising step, and maps them back to labels. Each question is a separate model request; SGLang can batch. Scores compare options, not calibrated probabilities.

Primitives:choicescorenoul

Technique worth stealing

Map options to letters, read answer-letter scores from the last active denoising step, and map back to labels.

Try it

Open http://localhost:8000/#doodle after launching with scripts/launch_diffusiongemma.py.

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-centricyes0.89
  • Shows a System One patternunclear0.42
  • Handles uncertaintyno0.06
  • Measuredno0.16
  • Runnableyes0.90
  • Worth recommendingno0.30
  • Model replicayes0.91
  • Problem scopescore on a 0–2 scale0.00
  • About Jevyes0.62

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 23 Sept 2026.