myJEV
Learning project that builds choose-an-answer decision models from scratch and from Qwen, with downloadable 0.8B and 4B releases.
Overview
myJEV is a learning and research project on how a model can make decisions inside an application. A request carries context, an instruction and the possible answers; the model scores the answers in one pass, and the application decides whether to accept the result or send it for review. It is inspired by TypeSafe's Jev but notes that Jev's architecture is undisclosed.
There are two starting points: a small network trained from random weights, and adaptations of pretrained Qwen models released on Hugging Face as myJEV-0.8B (CPU-friendly) and myJEV-4B (NVIDIA). The experiments compare what each learns, how well confidence identifies mistakes and what each costs to serve. A Docker image runs the myjev score command without a local Python install, and there are recorded demos and a CPU calibration lab. It is aimed at engineers who want to understand decision models hands-on.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which request-supplied answer is correct for this context and instruction? | — |
How it works
- Write a request with context, instruction and candidate answers.
- Score all answers in one pass with a myJEV model.
- Read the selected answer and its confidence.
- Accept above a threshold or route to review.
Strengths
- Runnable via Docker on CPU with no hosted endpoint.
- Compares from-scratch and pretrained approaches side by side.
- Focus on whether confidence flags mistakes.
Caveats
- Educational project rather than a production service.
- 0.8B CPU model is slower; 4B examples target NVIDIA GPUs.
Getting started
docker run --rm -v myjev-hf-cache:/cache/huggingface amitbahree/myjev:0.1.4-cpu myjev score --artifact bahree/myJEV-0.8B --device cpu --input /app/examples/request.jsonAt a glance
- Type
- Library
- Status
- research
- Language
- Python
- Runtime
- CLI, Docker
- Has code
- yes
Use cases and integrations
How curated lists describe it
Build decision models from scratch and fine-tune Qwen.
wh000wh000/awesome-jev-live
Frequently asked questions
What does myJEV return?
The selected answer from the options in the request plus its confidence.
Can I run it without a GPU?
Yes, the 0.8B release runs on CPU through the Docker image.
Is it a copy of Jev?
No; it is inspired by Jev's interface, while Jev's own architecture is undisclosed.
Related projects
Korean study notes and TypeScript mock-ups of Jev concepts, with a local Laya versus GPT-5 comparison and a calibration toolkit.
ProjectresearchTypeScriptPreprint survey of Jev-style typed decision models: their research roots, adaptation costs and what current evidence supports.
ModelresearchInteractive web workbench for exploring and benchmarking Jev's Choice, Noul and Score via OpenRouter, with threshold transfer curves.
ProjectprototypeTypeScript- cbjev★ 4
Self-hosted encoder model giving Jev-compatible choice, score and noul answers in one pass, pitched as a better-calibrated Laya successor.
ProjectresearchPython 33-chapter Chinese tutorial on agentic RL and GRPO, including a chapter on when Jev-style external decision models beat training with RL.
ProjectresearchPythonJoke decision model trained to answer wrongly with total confidence: 1.9% accuracy at 0.957 mean confidence, via a Jev-style API.
ProjectresearchPython