User repository on GitHub
Turning the Jev classifier model into an autoregressive next token predictor
- Forks
- 3
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Demos & Experiments
ChatJev
An experiment that drives Jev as an autoregressive next-token predictor: each step offers potential next tokens as choice options, and the resulting sequence of picks is the generated text. The author's verdict: it kinda works.
- Stars
- 7
- Language
- Python
- Reading time
- 2 min
- Published by
- Community
- Author
- erik-dunteman
- Last verified
- Oct 2026
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open source
Highlights
- Answers the claim that Jev cannot generate text by putting the classifier in an autoregressive loop.
- Each step gives Jev a list of potential next tokens as choice options and appends whichever one it picks.
- The README's honest summary is 'And it kinda works!' — it is a proof of concept, not a capable text generator.
- The repository is Python, with GitHub metadata showing it created and pushed on 20 September 2026.
Watch out
Grounded on the GitHub description, the madewithjev listing and a very short README reviewed 2026-10-15; there are no commands, results, dependencies or setup notes in the digest, so this entry stays deliberately thin. GitHub lists no licence for the repository.
Questions about this entry
Short answers drawn from the linked project or page.
A small experiment by erik-dunteman that drives Jev, a multiple-choice classifier, as an autoregressive next-token predictor: the model is repeatedly asked to pick the next token from a list of options, and the sequence of picks is the output text.
Each step presents the potential next tokens as choice options to Jev, which returns one typed answer. ChatJev appends the chosen token and loops, so text emerges from many single-token decisions instead of from a language model.
No. The README answers claims that Jev cannot generate text with 'FALSE!' and closes with 'And it kinda works!', which is the honest summary: it is a proof of concept, not a capable text generator, and the repository reports no metrics.
The digest does not include any install or usage instructions; the README is a short statement of intent. Setup notes, dependencies and which keys are required are not documented in the material this entry was built from.
It probes the boundary of the decision-model format. Reusing a choice question as a token loop is a reminder that Jev returns typed answers from declared options, and that the caller controls what each option maps to downstream.
It was listed by madewithjev as an experiment on the Jev classifier and targets Jev directly. The digest gives GitHub metadata — Python, stars 7, created and pushed 20 September 2026 — but no architecture detail beyond the loop description.
GitHub metadata lists no licence for the repository, and no paper or benchmark is referenced in the README, so treat it as a public experiment rather than a maintainable library.
More Jev resources like this
Text generation with a model that never generates text: jev-gpt asks Jev one typed question per choice — word type, WordNet category, a group of about 200 words split by embeddings, then the word — about 400 calls and 2 cents per prompt.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
florian-hoenicke/jev-gptPython
florian-hoenicke/jev-gptPythonA local app that turns a worded query into a physics pile of floating YC logos: retrieval over 6,241 companies by meaning, logo image, OCR'd logo text, and Jev re-applied tags, with Jev scoring the finalists into calibrated probabilities.
Jev Demos & Experiments#community#catalog-import#madewithjev
Community
aayan-dev/aayans-yc-indexorHTML
aayan-dev/aayans-yc-indexorHTMLA Next.js console whose whole design system is generated at runtime by Jev from a one-sentence brief: type 'an 18th century printed book' and every visual decision is chosen, not written, from a catalog a person authored.
Jev Demos & Experiments#community#catalog-import#madewithjev
Community
bilune/jev-designTypeScript
bilune/jev-designTypeScriptBrowser-only Laya demo suite: four Vite apps — a Traditional Chinese decision playground, Gomoku, Big Two and a 3D maze — run the Laya ONNX model on-device through Transformers.js, downloading fp16 weights from Hugging Face on first use.
Jev Demos & Experiments#community#catalog-import#madewithjev
Community
killkli/open-jev-layaPython
killkli/open-jev-layaPythonA Next.js and MUI showcase of Jev classifying Enron-Spam emails: one call asks a Noul spam question, a Choice category question and a Score urgency question at once. Includes a dataset view, a no-key example tab and a stand-alone Python script.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
harsha89/jev-demoTypeScript
harsha89/jev-demoTypeScriptCoaxes free text out of TypeSafe's classification endpoint by decoding one word at a time: every word is a separate choice question over a vocabulary, and the loop feeds its output back as the prefix. The README itself calls the approach a demonstration, not a good idea.
Jev Demos & Experiments#community#catalog-import#logicrw
Community
suidouble/let-jev-speakJavaScript
suidouble/let-jev-speakJavaScriptFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.