开发者在 MiniCPM5-2B 上构建结构化决策模型 MiniCPM5-2B-Jev
有人把 MiniCPM5-2B 改成了只输出概率的决策模型，JevBench 拿了 78.79%，做工单路由或规则核查的可以直接拿来用，代码全开源。
开发者 @ActbroDev 基于 OpenBMB 的 MiniCPM5-2B 改造出 MiniCPM5-2B-Jev，让它输出预定义选项的概率而非自由文本，可用于工单分类、条件核查和评分等场景。项目在 JevBench 上取得 78.79%，在 10-source decision-v7 基准上准确率达 86.41%，ECE 为 0.0307。模型提供 /v1/systemone 结构化决策 API，支持多问题共享 KV-cache，代码、LoRA 适配器和校准头均以 Apache 2.0 开源。
What if your model could make a structured choice instead of writing another paragraph? 🎯
Developer @ActbroDev built MiniCPM5-2B-Jev on OpenBMB’s MiniCPM5-2B, adapting it into a System 1 decision model.
📄 Give it a document and a set of questions. It returns probabilities for predefined options, without generating a free-form answer—helping developers: - Classify requests and route support tickets - Check whether conditions or requirements are met - Score cases against defined criteria
📊 Results reported by the project: 🏆 78.79% on JevBench, ranking #1 among ≤2B open-weight System 1 / Jev models ✅ 86.41% accuracy on the 10-source decision-v7 benchmark 📏 0.0307 ECE on decision-v7, measuring how closely confidence matches accuracy
🛠️ Built for developers: • Structured decision API via /v1/systemone • Shared-context KV-cache reuse across multiple questions • Open-source code, LoRA adapters, and calibration head under Apache 2.0 • Training and evaluation scripts to explore and reproduce the approach
From support-ticket routing to criteria-based scoring, this project shows how the community is adapting MiniCPM for practical decision workflows. 🌍
🙌 Shout-out to yuting-ai for building and sharing!
🔗 Explore Project: https://t.co/VRkkO4cY7G 🤗 Get MiniCPM5-2B: https://t.co/TNUu629lXn
#MiniCPM #OpenBMB #OpenSourceAI
