--- title: JEV-Demo canonical_url: "https://www.modelscope.ai/studios/AutoTrust-Labs/JEV-Demo" md_url: "https://www.modelscope.ai/studios/AutoTrust-Labs/JEV-Demo.md" repository: AutoTrust-Labs/JEV-Demo chinese_name: "JEV Interactive Demo" last_updated: 2026-10-16 sdk_type: docker sdk_version: 6.29.1 downloads: 0 stars: 1 --- # JEV-Demo > JEV-Demo - An interactive demo by AutoTrust-Labs on ModelScope. Explore AutoTrust JEV decision models in English: text and image decisions, confidence-aware reasoning, game arenas, news search, and short-video recommendations. AutoTrust-Labs/JEV-Demo is an interactive demo (Studio) on ModelScope. built with docker 6.29.1. also known as 「JEV Interactive Demo」. - **Repository**: AutoTrust-Labs/JEV-Demo - **SDK**: docker - **SDK version**: 6.29.1 - **Downloads**: 0 - **Stars**: 1 - **Last updated**: 2026-10-16 Source: https://www.modelscope.ai/studios/AutoTrust-Labs/JEV-Demo --- # JEV Interactive Demo An English interface for exploring AutoTrust's JEV models and applications. It brings together material from all five public Hugging Face Spaces: JEV-27B-Demo, jev-9b-decision-demo, news, shorts, and README. ## Features - Decision Lab: text and image input for yes/no judgments, multiple-choice decisions, and 0–5 scores, with model probabilities, latency, and raw responses. Upload, drop, or paste one static PNG, JPG, or WebP image up to 5 MB, preview it, then enter a question and options. Text context is optional; results indicate whether an image was included. - Confidence-based reasoning: `thinking=off`, `auto`, or `on`, with a confidence threshold and reasoning budget, using the same JEV-27B-VL service. The native protocol supports only `off` for scoring tasks. - Game Arena: Mario, Rubik's Cube, Tetris, and Doom. The first three connect to the original online service through an English interface. - Doom: ViZDoom / Freedoom runs on this Space's CPU and sends the current screenshot and structured state to JEV-27B-VL. Up to four sessions, ten steps per game, and cleanup after five minutes of inactivity. - News search: current stories, model ranking versus the search engine's original order, and answers with sources. - Short-video recommendations: interest selection, search, probabilities, feedback, explanations, and preferences. - Recorded demos, model information, source links, and license notices. ## Deployment Deploy as a Docker Space on ModelScope's international site, under the `AutoTrust-Labs` organization. The application listens on port **7860** and does not load model weights. The Docker image uses the official Python base image. Choose an available CPU resource configuration supported by the destination platform. The complete source uses a standard directory layout. The Dockerfile copies `static/` directly into the image. Configure these **secret environment variables** in the Space: - `JEV_BASE_URL`: the native JEV API base URL, such as `https://api.example.com`, without a trailing `/v1`. - `JEV_API_KEY`: an existing API key for that service. - `JEV_MODEL`: `autotrust/JEV-27B-VL`. The API key is used only on the server. Do not place it in HTML, JavaScript, a public README, or image build arguments. Without a configured URL and key, the Decision Lab and Doom AI explicitly report that they are not ready; they do not substitute another model. Native requests use `POST /v1/decide`; image input uses `state: [text, {"image": dataURL}]`. The API supports `thinking`, `threshold`, `think_budget`, and `return_reasoning`. Responses preserve `probabilities`, `choice_index`, `elapsed_seconds`, and `thinking`. ## Original online services News and short videos connect to `news.guru.so` and `shorts.guru.so`. The three-game arena connects to fixed public API routes on `autotrust-jev-27b-demo.hf.space` and also depends on the original project's game worker. Availability and model configuration are managed by those upstream services; those features are not direct inference through this Space's configured JEV-27B-VL API. News answers, retrieval, and related operations may use other models managed by the original services. All direct inference added by this application uses the configured JEV-27B-VL API, with no additional model deployment required. Initial previews, recordings, and cold-start recommendations are labeled separately from live inference. Visitor submissions are sent to the relevant service. This Space does not write decision inputs to a database. The anonymous short-video user ID is stored in the current browser's `localStorage`; interests and playback feedback are stored by the upstream demo service, with a clearing option in the interface. The application does not read browser account credentials. ## Local development and checks Use Python 3.12: ```sh pip install -r requirements.txt python app.py ``` Open `http://127.0.0.1:7860`. ViZDoom is installed on Linux; it is omitted by default on other platforms, without affecting the other pages. Run the backend checks with: ```sh python -m pytest -q tests ``` The interface and documentation are in English. News, videos, and model-generated text may retain their source language. See `static/NOTICE.txt`, `GAMES_NOTICES.md`, and `UPSTREAM_NOTICES.md` for provenance and third-party notices.