Home /Open models and replicas /metask-jev
metask-jev
Open-weight typed-decision models on Qwen3.5 with calibrated per-option probabilities in one forward pass; reports 80.1 percent on JevBench against 75.3 for Jev 1.13.
From the readme
Metask Jev-Lab 探索最优开放权重的 typed-decision（Jev-class）模型：给定 state + bounded rubric，单次前向输出每个选项的校准概率。无生成、无解析。 模型一览 模型 基座 13 子集宏平均 JevBench 231 权重 ---------:---:--- metask-jev-4b Qwen3.5-4B 79.6% 80.1% (@4096) HF · 魔搭待发 metask-jev-0.8b Qwen3.5-0.8B 74.7% 55.8% 未发布 参照：Bespoke Nimble-9B 74.8% / 63.5（第14名）；Jev 1.13.0 76.0% / 75.3（第2名）。 一键复现推理 输入一段 state + …
Details
- Section
- Open models and replicas
- owner
- metask-ai
- stars
- 1
- forks
- 3
- pushed
- 2026-09-22
- Language
- Python
What it is
- Form
- Library or SDK
- Host agent
- Standalone
- Audience
- Researchers
Maturity
Docs
reports measured numbers