What does the demo accept as input?

Typed questions over text and images: noul as a yes/no statement answered P(true), choice among 2-16 options you list, or score on a fixed 0-5 scale, from one prefill pass.

Is the model trained on images?

No. The card says the decision head was trained on text only, so every decision over an image is zero-shot — ranking covers, judging screenshots, classifying photos.

How does the demo run on ZeroGPU?

It ports the authors' vLLM reference server to plain transformers plus peft, renders the raw bare-v1 prompt with vision placeholders, and loads weights in 4-bit NF4, keeping lm_head in bf16.

What is System 2 for?

System 2 is the untouched Qwen3.8-27B with the LoRA disabled, available for deliberate step-by-step reasoning, including over images, when the one-pass answer is not enough.

Which images are used in the examples?

CC0 stock photos from linoyts/repo-to-space-example-inputs, with question phrasings following the model card and the authors' JEV-27B-DEMO repository under Apache-2.0.

Why is lm_head kept in bf16?

Because the decision head is a LoRA on lm_head, so its precision is System 1's precision; the rest of the weights load in 4-bit NF4 through bitsandbytes.

Is this the authors' reference implementation?

No. It is a demo port of the reference server to transformers and peft so it can run on ZeroGPU; inference follows the same bare-v1 prompt and calibration temperatures but is not the vLLM deployment.

CommunityJev Demos & Experiments

JEV-27B-VL Demo

A ZeroGPU demo of AutoTrust's JEV-27B-VL: System 1 returns calibrated noul, choice, and score probabilities from one prefill pass over text and images, with System 2 on the untouched Qwen3.8-27B for escalation.

See highlights

Open huggingface.co

Format: Playground
Published by: Community
Author: hugging-apps
Last verified: Oct 2026

Tagscommunity playground demo gradio multimodal

Works forJev for Classification Jev for Scoring Jev for Decision gating

Used inJev for Software & DevTools Jev for Media & Entertainment

Languages: Python
Platforms: Browser
Capabilities: Open sourceRuns locally

Highlights

System 1 answers noul as P(true), choice over 2-16 options, and a fixed 0-5 score in a single prefill pass with a calibrated distribution.
The decision head was trained on text only, so image decisions are zero-shot: the card names ranking covers, judging screenshots, and classifying photos.
Inference ports the authors' vLLM reference server to transformers plus peft for ZeroGPU, using the raw bare-v1 prompt with vision placeholders.
Weights load in 4-bit NF4 with lm_head kept in bf16 — the decision head is a LoRA on lm_head — and ZeroGPU streams them into VRAM per request.
System 2 is the untouched Qwen3.8-27B with the adapter disabled, available for step-by-step reasoning, including over images.

Watch out

Built on autotrust/JEV-27B-VL under Apache-2.0; the README states the decision head saw text only, so image behavior is zero-shot rather than a trained skill. It is a 4-bit NF4 port to ZeroGPU for the demo, not the authors' vLLM server, and weights are streamed per request rather than kept resident.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

LinkedIn: What multimodal decision models actually change, and what they do not.

LinkedInJevDirectory.org

Related terms

Glossary definitions related to this entry.

Playground

More Jev resources like this

JEV-9B Decision Demo

An interactive ZeroGPU demo of AutoTrust's JEV-9B: System 1 answers noul, choice, and score questions in one prefill pass with calibrated probabilities, while an Escalate tab hands low-confidence cases to the untouched Qwen3.5-9B base.