jev-lm
Word-level language model whose output layer is Jev: next-word choices, Noul-verified drafted chunks and bits-per-token evaluation.
Overview
Jev-LM is a word-level language model whose output layer is Jev. Because Jev never emits text, everything an LLM normally hides lives in the code: the tokenizer, sampler, repetition mask, stop rule and a verified-chunk path standing in for a KV cache. An n-gram drafter proposes continuation chunks that Jev verifies with Noul questions, so several words can be accepted per round trip; the README's example generates nine words in two calls.
Commands probe the next-word distribution, verify candidate continuations, generate text, and evaluate bits per token against a unigram baseline. Answers are cached by a hash of state, questions and model so re-running an evaluation costs nothing. It has zero dependencies, runs on Node 18+, works as a CLI or library, and reports measured results against jev-1.13.0 in the README.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which word comes next? | — |
| Yes / no | Is this drafted chunk a good continuation? | — |
How it works
- Drafter proposes a continuation chunk
- Jev verifies the chunk with Noul or chooses the next word
- Apply repetition mask and stop rule
- Cache answers and evaluate bits per token
Strengths
- Exposes all LM machinery in readable code
- Chunk verification reduces round trips
- Caching makes evaluations repeatable
Caveats
- Research experiment, not a practical generator
- Many calls for long text
Getting started
`npx jev-lm gen "she went to the door" --draft --trace` with TYPESAFE_API_KEY set.At a glance
- Type
- Project
- Status
- research
- Language
- TypeScript
- Runtime
- Node 18+ CLI and library
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
A word-level generation experiment that asks Jev to select words or verify locally drafted continuations.
logicrw/awesome-jev-projectsA word-level language model whose output layer is Jev: n-gram drafter, Noul…
wh000wh000/awesome-jev-liveA word-level language model whose output layer is Jev: n-gram drafter, Noul chunk verification, bits-per-token eval.
hellogumbo/awesome-jev
Frequently asked questions
How does Jev generate text here?
By choosing the next word or verifying drafted chunks.
How is it evaluated?
Bits per token against a unigram baseline.
Does it have dependencies?
No.
Related projects
- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython Research decoder that treats Jev as a language model, assembling replies from repeated choices over phrases, words and letters.
ProjectresearchPythonReads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.
ProjectresearchPythonRunnable yardstick that tests Jev's three headline claims on a small local model: typed output, no hallucination, calibration.
ProjectresearchPythonOpen Jev-class decision models on Qwen3.5 with candidate-logit readout; the 4B model reports 80.1% on JevBench's 231 items.
ProjectresearchPythonOpen Jev replica on Gemma 4 E2B and Gemma 3 270M served on Modal, with eight demos and a TypeSafe eval head-to-head.
ProjectresearchPython