Open Jev (intikhab49)
Open 150M encoder that answers noul, choice and score questions in one pass, trainable on a free Colab T4 in about 30 minutes.
Overview
This Open Jev reproduces Jev's interface with an open 150M encoder that answers arbitrary typed questions about a state in one non-autoregressive pass with calibrated confidence. It supports noul (boolean), choice (enum with descriptions) and score (ordered scale), and questions are supplied per request rather than baked into weights. It can be trained on a free Colab T4 in under 30 minutes and run locally or in a browser.
The README reports it is 0.03 behind Jev on its benchmark (0.697 vs 0.727 per the list summary), 2.5x better calibrated and 4x faster, at no cost. It notes TypeSafe does not publish Jev's parameter count, so size comparisons are inference. A GitHub Action is offered for typed decisions in CI, and the example asks an observability system whether to continue, queue for human review or stop an agent, plus a risk score. It is for developers wanting a cheap, trainable decision model.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What should the observability system do? | continue · human_review · stop |
| Score | How risky was this behaviour? | Benign · Low · Moderate · High |
How it works
- Train the 150M encoder with the Colab notebook.
- Call decide(state, questions) locally or in the GitHub Action.
- Read labels, confidences and probabilities.
- Escalate when confidence is low.
Strengths
- Free to train and run.
- Reported better calibration than Jev on its benchmark.
- GitHub Action for CI decisions.
Caveats
- Slightly lower accuracy than Jev per the README.
- Benchmark comparison relies on the LocalLLaMA typed-decisions dataset.
Getting started
Open jevlite_colab.ipynb in Colab to train, then call decide() with Question.choice / Question.score / noul questions.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch, browser, GitHub Action
- Has code
- yes
Use cases and integrations
How curated lists describe it
Open reproduction of TypeSafe Jev: a 150M typed decision engine…
wh000wh000/awesome-jev-liveOpen reproduction of TypeSafe Jev: a 150M typed decision engine (noul/choice/score in one non-autoregressive pass, calibrated confidence). 0.697 vs Jev's 0.727, 2.5x better calibrated, 4x faster, free. Trains on a Colab T4 in 30 min.
hellogumbo/awesome-jevA 150M encoder trained to answer typed questions in one pass, with the training notebook written to run on a free GPU.
MrJev/awesome-jev
Frequently asked questions
How big is the model?
150M parameters.
How does it compare to Jev?
About 0.03 lower accuracy on its benchmark, with better calibration and 4x speed per the README.
How long does training take?
Under 30 minutes on a free Colab T4.
Related projects
- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython 33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonPre-registered field test of Jev, GLiNER2.5-Decide, Laya and other decision models against small LLMs on fresh September 2026 data.
ProjectresearchPythonReads Jev-style choice, yes/no and score probabilities from masked diffusion language models with shared Yes/No weights.
ProjectresearchPythonOpen Jev-class decision models on Qwen3.5 with candidate-logit readout; the 4B model reports 80.1% on JevBench's 231 items.
ProjectresearchPythonOpen System-1 (OpenJev): Apache-2.0 research models giving calibrated per-option probabilities, best is a 4.9B exp9a.
ProjectresearchPython