autotrust/JEV-27B-Demo

Space

JEV-27B Live Demo

29

23 commits

1 linked in READMEs

updated Sep 30, 2026

See the code

README

JEV-27B Live Demo

AutoTrust · Model

Launch a live JEV-27B game run in the game arena.

  • Mario: original NES World 1-1, with movement and jump decisions.
  • 3D Rubik's Cube: a 25-turn scramble; select a seed or create a new scramble.
  • Tetris: smooth gravity acceleration, continuing at maximum speed until 20 lines.
  • The right panel shows the selected action, option probabilities, and current / average individual model inference duration. Mario movement and jump are separate calls.
  • Start and stop runs yourself; one run per game executes at a time, with a short queue. Recent runs reconnect when you reload the page.

The game worker runs on AutoTrust's existing B300 deployment. Model inputs are structured observations from the game engines. The display label is jev-27b-vl; the inference checkpoint is autotrust/JEV-27B. Cube actions use a fixed library of standard algorithms. Cube turns are animated independently of model inference timing. All raw run traces and rendered game frames are retained on the B300. The Space contains no Mario ROM.

The decision lab and Doom arena and demo reel remain available.

  • Decision lab: text or image input for multiple choice, true/false, and 0–5 scoring, with option probabilities and timing. Upload one JPEG, PNG, or WebP image up to 5 MB, add a question, and optionally add text context.
  • Doom arena: target practice and center defense. Play manually, request one JEV move, or run ten consecutive model decisions.
  • Demo reel: ten recorded demonstrations and six benchmark comparison bar charts.

Live decisions use the existing AutoTrust JEV-27B deployment. Image input uses Qwen/Qwen2-VL-2B-Instruct, pinned to revision 895c3a49bc3fa70a340399125c650a463535e71c, to describe the image before JEV makes the decision. This is a vision-assisted pipeline, since the deployed JEV checkpoint is text-only. The interface names the vision model and shows image-understanding time separately from JEV decision time. Images are processed in memory, normalized to at most 1024 pixels per side, and are not saved by this app. The Space runs the interface and the ViZDoom simulator on CPU. In Doom, JEV receives structured engine state and the simulation pauses between decisions. Game sessions expire after ten minutes of inactivity; up to eight sessions and four simultaneous model calls are supported.

The deployed inference revision is 7ce153d6f7933769527939bcb92ec3f3d8292884. The recorded reel is the latest ten-demo 5× edit (41 seconds), with the September 27, 2026 benchmark comparison. Benchmark details are in benchmark-notes.md.

Deployment

Run as a Docker Space on CPU Basic, port 7860. Configure these values in Space settings:

  • JEV_BASE_URL: HTTPS origin of a compatible backend providing /decide and /v1/models.
  • JEV_API_KEY: secret bearer token for that backend. Keep it in Space Secrets.
  • Optional HF_GAMES_WORKER_TOKEN: a separate worker authentication secret. By default, the app derives a domain-separated worker key from the existing server-side JEV_API_KEY; no additional Space permission or secret is required. The worker receives only this derived key, never the browser.

The browser calls only this Space's API. Backend credentials remain on the server. Model requests have bounded input sizes and concurrency. Errors are shown directly; no substitute model or scripted AI controller is used.

Credits

Doom scenarios run using ViZDoom and its bundled Freedoom assets.

Reel soundtrack: “Tech Live” by Kevin MacLeod (incompetech.com), licensed under Creative Commons Attribution 4.0. Edited excerpt.

See THIRD_PARTY_NOTICES.md for component and media attribution.

autotrust
decision-making
docker
games
vizdoom

autotrust/JEV-27B-Demo

Space

JEV-27B Live Demo

29

23 commits

1 linked in READMEs

updated Sep 30, 2026

See the code

README

JEV-27B Live Demo

AutoTrust · Model

Launch a live JEV-27B game run in the game arena.

  • Mario: original NES World 1-1, with movement and jump decisions.
  • 3D Rubik's Cube: a 25-turn scramble; select a seed or create a new scramble.
  • Tetris: smooth gravity acceleration, continuing at maximum speed until 20 lines.
  • The right panel shows the selected action, option probabilities, and current / average individual model inference duration. Mario movement and jump are separate calls.
  • Start and stop runs yourself; one run per game executes at a time, with a short queue. Recent runs reconnect when you reload the page.

The game worker runs on AutoTrust's existing B300 deployment. Model inputs are structured observations from the game engines. The display label is jev-27b-vl; the inference checkpoint is autotrust/JEV-27B. Cube actions use a fixed library of standard algorithms. Cube turns are animated independently of model inference timing. All raw run traces and rendered game frames are retained on the B300. The Space contains no Mario ROM.

The decision lab and Doom arena and demo reel remain available.

  • Decision lab: text or image input for multiple choice, true/false, and 0–5 scoring, with option probabilities and timing. Upload one JPEG, PNG, or WebP image up to 5 MB, add a question, and optionally add text context.
  • Doom arena: target practice and center defense. Play manually, request one JEV move, or run ten consecutive model decisions.
  • Demo reel: ten recorded demonstrations and six benchmark comparison bar charts.

Live decisions use the existing AutoTrust JEV-27B deployment. Image input uses Qwen/Qwen2-VL-2B-Instruct, pinned to revision 895c3a49bc3fa70a340399125c650a463535e71c, to describe the image before JEV makes the decision. This is a vision-assisted pipeline, since the deployed JEV checkpoint is text-only. The interface names the vision model and shows image-understanding time separately from JEV decision time. Images are processed in memory, normalized to at most 1024 pixels per side, and are not saved by this app. The Space runs the interface and the ViZDoom simulator on CPU. In Doom, JEV receives structured engine state and the simulation pauses between decisions. Game sessions expire after ten minutes of inactivity; up to eight sessions and four simultaneous model calls are supported.

The deployed inference revision is 7ce153d6f7933769527939bcb92ec3f3d8292884. The recorded reel is the latest ten-demo 5× edit (41 seconds), with the September 27, 2026 benchmark comparison. Benchmark details are in benchmark-notes.md.

Deployment

Run as a Docker Space on CPU Basic, port 7860. Configure these values in Space settings:

  • JEV_BASE_URL: HTTPS origin of a compatible backend providing /decide and /v1/models.
  • JEV_API_KEY: secret bearer token for that backend. Keep it in Space Secrets.
  • Optional HF_GAMES_WORKER_TOKEN: a separate worker authentication secret. By default, the app derives a domain-separated worker key from the existing server-side JEV_API_KEY; no additional Space permission or secret is required. The worker receives only this derived key, never the browser.

The browser calls only this Space's API. Backend credentials remain on the server. Model requests have bounded input sizes and concurrency. Errors are shown directly; no substitute model or scripted AI controller is used.

Credits

Doom scenarios run using ViZDoom and its bundled Freedoom assets.

Reel soundtrack: “Tech Live” by Kevin MacLeod (incompetech.com), licensed under Creative Commons Attribution 4.0. Edited excerpt.

See THIRD_PARTY_NOTICES.md for component and media attribution.

autotrust
decision-making
docker
games
vizdoom