Are the benchmark numbers real?

They are the author's controlled comparison: identical data, splits, and hand-written holdout for both models, with full per-question rows shipped in results files. Note the two Instruct seeds disagree on 11 questions, so near ties are noise.

Why is it slower than MiniJev if it scores higher?

Each decision scores one full text path per candidate through a 0.6B backbone, measured near 1.4 seconds per case against 1.6 milliseconds for MiniJev. Quality versus speed is the stated engineering trade.

Can one port serve decisions and generation?

Yes. The fused server routes decisions with the adapter on and completions or chat with it off, serializing inference under a lock. A decision issued mid-stream returns bit-identical results with no state leakage.

What hardware does this assume?

Apple Silicon with the Metal backend; full fine-tuning fits but swaps under pressure, so the documented path is frozen-backbone LoRA. Backbone weights download from Hugging Face on first run.

Do I have to train it before I can use it?

No. A pretrained rank-16 adapter (about 40MB, seed 17) ships in the repository so you can serve and evaluate without training; the Qwen3-0.6B backbone downloads from Hugging Face on first run.

How do the Base and Instruct variants compare?

JevNext-Base scored 91.2% on the 102-question holdout against Instruct's 86.3% and 87.3% across two seeds, but a paired McNemar analysis puts them at p about 0.27, so the README calls the two backbones statistically indistinguishable for this task.

Is the training data real or synthetic?

Synthetic and rule-generated, shared verbatim with MiniJev so results compare across projects; the holdout is hand-written, and the README warns that model behavior on real-world text is not guaranteed.

CommunityJev Models & Reimplementations

JevNext

Learning project rebuilding a Jev-style decision model on a frozen Qwen3-0.6B backbone with a rank-16 LoRA, served alongside text generation from one weight set on Apple Silicon. Explicitly not production software.

See highlights

Open github.com

Stars: 2
Language: Python
License: MIT
Published by: Community
Author: DejaAI2
Last verified: Oct 2026

DejaAI2/JevNext

User repository on GitHub

View on GitHub

Decision model on a Qwen3-0.6B backbone with LoRA: fused decision endpoint (/v1/systemone) + OpenAI-compatible chat completions with thinking, SSE streaming and vLLM-style sampling, running on Apple Silicon MPS.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

MiniJev, minus the ceiling: a Jev-style decision model on a Qwen3-0.6B backbone (LoRA, rank 16), fused with a full OpenAI-compatible generation server — one set of weights, two capabilities, zero-copy LoRA switching. Built and benchmarked end-to-end on Apple Silicon (MPS).

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

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Same MiniJev data and holdout protocol; reported accuracy rises from 58.8% into the high eighties.
One server exposes decision scoring plus OpenAI-style completions and chat, flipped by a zero-copy adapter flag.
Decision route is non-autoregressive candidate scoring with zero output tokens; generation flips adapter off.
Ships a 40MB pretrained adapter so you can serve and evaluate without the full training run.

Watch out

Educational synthetic-data model with real per-decision latency costs; the author forbids using it for real routing or triage without human oversight. 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.

More Jev resources like this

gut

Python helper that turns judgment calls into one-liners like likely, classify and rate, defaulting to Jev while also running on local or OpenAI-compatible backends.