Does Jev beat a simple baseline as a text generator?

It does not. Bits per token on the shipped held-out file favour the unigram table over Jev, and character granularity is worse. The digest frames the value as exposing what LLMs normally hide, not as a shippable generator.

Why is chunk verification the performance trick here?

One request per step carrying a 229-option Choice plus Noul checks, at about a quarter second median per round trip. Adding chunk-verification questions costs no extra wall clock, while spelling word by word is far more expensive.

What breaks without the masking and stop rules?

Greedy output loops on repeats, the end option fires on noise, and a confident done signal would stop zero-length prompts. The mask plus min_new and dual-signal agreement rules exist to counter exactly those measurements.

Which commands and setup do I use to reproduce it?

Probe distributions, verify candidate continuations, generate with tracing, and eval bits per token against the baseline. Answers cache by state hash, and the key comes from environment, key file, or token file.

Why build a language model on a decision model at all?

It is a deliberate teaching exercise. The README's stated reason is the inverse of writing a wrapper: everything an LLM normally hides, the tokenizer, sampler, repetition mask, stop rule and a stand-in for the KV cache, is code you have to write by hand here. The point is not to ship a generator, and it says so.

How large can the vocabulary be?

The shipped vocabulary is 229 options. A Choice accepts up to 255 and TypeSafe documents reliable behaviour up to roughly 240, so the flag will take a slightly larger file than the default but not a much larger one. That ceiling is why the whole model is word-level rather than subword or character.

How does caching change what a run costs?

Re-running is free when nothing changed. Answers are cached under XDG_CACHE_HOME or ~/.cache/jev-lm, keyed by a hash of the state, the questions and the model, so repeating an eval costs nothing; a no-cache flag turns that off when you want fresh distributions.

CommunityJev Practices & Patterns

Jev Lm — y0usaf

Word-level language model experiment that uses Jev as its output layer, with explicit tokenizer, sampler, repetition mask, stop rules, and chunk verification.

See highlights

Open github.com

Stars: 7
Language: TypeScript
License: MIT
Published by: Community
Author: y0usaf
Last verified: Oct 2026

y0usaf/jev-lm

User repository on GitHub

View on GitHub

A word-level language model whose output layer is Jev: n-gram drafter, Noul chunk verification, bits-per-token eval

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

A word-level language model whose output layer is Jev. Jev never emits text, so everything an LLM normally hides lives in this code: the tokenizer, the sampler, the repetition mask, the stop rule, and a verified-chunk path that stands in for a KV cache. One HTTP request per round trip, zero dependencies, Node 18 or newer. Set a key one of three ways: TYPESAFEAPIKEY, a file named by TYPESAFEAPIKEYFILE, or /Tokens/TYPESAFEAPIKEY.txt.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification Jev for Evaluation

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Generates text one Jev request per step over a 229-word vocabulary with caching.
Provides probe, verify, gen, and eval commands plus a library API with drafter support.
Publishes round-trip latency, token costs, and bits-per-token comparisons with a baseline.
Documents masking, dual stop signals, tolerance-based chunk keeping, and sparse probabilities.

Watch out

Word path costs thousands of tokens per word and measured quality trails a unigram table, so treat it as an experiment. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

Other listings of this project

Forks and copies of this project in the directory. This entry is the canonical one.

jev-lm

A character-level language model assembled out of a classifier: the vocabulary becomes 28 Choice options, a to z plus space and stopping, and generation is the ordinary autoregressive loop.

Jev Tools & CLIs#community#awesomejev#catalog-import

Community

uhhfeef/jev-lmPython

More Jev resources like this

jev-lab

A small Node.js TypeScript CLI for trying Jev on synthetic support text: one urgency question, three questions in a single request, or a latency comparison. No web server and no UI.