Skip to content
プロジェクト

denis-pplx/autojev has 134 GitHub stars and 15 forks. GitHot measured 0 stars in 24 hours and 0 in 7 days. Last checked 2026-11-20 00:14 UTC. Last push 2026-09-22.

この数値の測り方

計測したスター増加

24時間
0
7日間
0
15日間
履歴を収集中
1か月
履歴を収集中

期間ごとの純増数です。期間は重複しており、日別の値ではありません。未計測の値は空欄です。スター数は注目度の指標であり、品質の保証ではありません。

概要

Full-weight Qwen decision model with calibrated probabilities, training code, REST API, and playground.

機械学習

AutoJev-27B

Built with autonomous agents, from research and data generation to training, evaluation, and deployment. The user set the goals and refined the scope; agents executed the work.

AutoJev-27B is a multimodal decision model based on Qwen3.8-27B. It returns probabilities over supplied choices in one forward pass per question, with a TypeSafe-compatible API and browser playground.

Code · Model weights

Results

Benchmark accuracy: Qwen3.8-27B, AutoJev-27B and Jev

Model Overall accuracy ↑ ECE ↓ Brier ↓
Qwen3.8-27B 69.83% 0.06483 0.40834
AutoJev-27B 84.60% 0.04282 0.22027
Jev 82.79% 0.05274 0.25400

Training

One H200 · full-weight SFT · 73,000 unique training examples · 286 updates. The released model is checkpoint 200. Training uses cross-entropy; calibration fits a scalar temperature separately.

Training loss

Run bash configs/train.sh --help for training arguments. The exact curated training corpus is not bundled.

Run

Python 3.12+, uv and a GPU with space for approximately 49 GiB of BF16 weights plus runtime overhead.

git clone https://github.com/denis-pplx/autojev.git
cd autojev
uv sync --frozen --python 3.12
uv run hf download denis-pplx/autojev-27b --local-dir checkpoints/selected
AUTOJEV_CHECKPOINT=checkpoints/selected uv run autojev-serve

Open http://localhost:8000 for the playground or /docs for the API. POST /v1/systemone supports choice, noul, score, and optional base64 images. Set AUTOJEV_API_KEY to enable authentication. While the weights are private, authenticate with uv run hf auth login before downloading.

Use the included DecisionModel loader or server. Published benchmarks measure text decisions; image support is not a natural-image accuracy claim.

Code: MIT · Weights: Apache 2.0 · Independent implementation inspired by Jev.