User repository on GitHub
an open-source, from-first-principles reconstruction of the ideas behind TypeSafe AI's Jev, written in pytorch
- Forks
- 19
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Open Jev — kyegomez
Unofficial PyTorch reconstruction of the ideas behind Jev, encoding state once and answering each question through small typed readout heads in a single parallel pass.
- Stars
- 99
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- kyegomez
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is Open Jev — nico-martin.
About this project
An open-source, from-first-principles reconstruction of the ideas behind TypeSafe AI's Jev, written in PyTorch. Jev is presented as a system one model: unstructured program state goes in, typed probabilistic decisions come out. Instead of generating text one token at a time, this implementation encodes the state once and answers every question through small typed readout heads.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Implements Noul probability, Choice over runtime-supplied options, and Score with ordinal thresholds plus expectation.
- Questions are isolated in the batch dimension and cross-attend to a shared cacheable state representation.
- Training loss accepts soft probability targets with calibration and consistency terms instead of one-hot labels.
- Ships with a deterministic hash tokenizer adequate for running the architecture but not for training a useful model.
Watch out
Weights are random with no trained checkpoints, so outputs are illustrative and calibration must be remeasured before unattended use. 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.
There are no matching performance claims here: the author states this does not reproduce TypeSafe training data or weights and makes no claim of matching published results.
Only structural validity, such as a Choice answer being one of the declared options; it does not guarantee the prediction is correct under distribution shift.
Install PyTorch, construct the model from a small config, pass a state dict with Noul, Choice, and Score questions, and run the module script for a calibration-oriented training step.
A trained tokenizer, a pretraining and distillation pipeline, scaling benchmarks, calibration evaluation, and published checkpoints are all listed as future work.
It encodes the state once with a bidirectional transformer, lets learned query slots cross-attend into one shared cacheable representation, then small typed heads emit noul, choice or score outputs in a single parallel pass.
RLCDLoss combines soft-target NLL, Brier score, consistency, confidence and ECE terms, and expects probability distributions rather than one-hot labels so disagreement and ambiguity stay visible.
Replace the hash tokenizer with a trained one, add a real pretraining and distillation pipeline, benchmark state-cache reuse and multi-question scaling, evaluate calibration and distribution shift, and publish checkpoints.
More Jev resources like this
46GitHub stars
Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.
Jev Demos & Experiments#community#jevbest#catalog-import
941GitHub stars
Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
TypeLLM/TypeLLMPython
TypeLLM/TypeLLMPython55GitHub stars
Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
fidecastro/jevifyPython
fidecastro/jevifyPython164GitHub stars
End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zwliJay/jev-forgePython
zwliJay/jev-forgePython70GitHub stars
Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
bladedevoff/stuntdPython
bladedevoff/stuntdPython42GitHub stars
Open harness that renders allowed JSON answers on a diffusion canvas, denoises for a fixed step budget, then picks the top allowed tokens from final logits with no training.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
JoshuaSP/open-jevPython
JoshuaSP/open-jevPythonFrom the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.