datawhalechina/jev-cookbook
⚡ 适合中国宝宝的 Jev 入门教程｜手把手带你了解关于 Jev 的一切——Jupyter Notebook 轻松实验，从三种问题原语到 18 篇实战配方、语音智能家居、模型评测与本地微调，全面掌握 System One 判断模型的开发范式
What it solves
大模型应用中的文本解析环节缓慢、昂贵且脆弱。Jev Cookbook提供了一条新路线：使用System One模型原生输出类型化决策与校准概率，避免解析。中文教程此前空白，本项目填补了这一缺口。
How it works
项目包含十一章可运行的中文教程，使用Jupyter Notebook。用户通过发送状态与类型化问题（Choice/Score/Noul三种原语）获取结构化答案。教程从入门到实战，覆盖18篇官方配方、语音智能家居实验、模型评测（231道公开题四维对比）、以及用RLCD微调开源同类模型Laya的完整流程。所有概率、延迟、token与成本均来自可复核的实测数据。
Who it’s for
面向具备基础Python能力、想入门“AI对软件”开发范式的开发者、工程师与学生，不需要机器学习背景。前置要求是Python≥3.10和Jupyter Notebook，API Key可选（免费注册）。
Highlights
- 零门槛入门：不配置API Key也能用离线示例学习全部代码路径
- 三种原语问答纪律：Choice/Score/Noul，一题只问一件事
- 18篇官方实战配方：重排序、语义搜索、结构恢复、函数调用等，每篇附工程坑
- 完整应用实战：语音驱动3D智能家居、12个可运行游戏与应用项目
- 诚实评测纪律：231道公开题四维对比，无重试、无回退、预算记账
- 本地微调全流程：用RLCD微调开源模型Laya，构建中文数据集并部署本地服务
- Agent集成范式：在Pi与DSH框架中集成判断门控
Related
- Project
apple/foundation-models-utilitiesA utility package for the Apple Foundation Models framework that provides chat completions clients, context window management, and dynamic skill activation for LLMs.
- Project
ZLKong/Awesome-Collection-Token-ReductionA curated, up-to-date collection of research papers on token reduction (pruning, merging, compression) for generative AI, organized by year and application area, with a companion position paper and structured Google Sheet index.
- Project
SilentCoderHere/aihubA privacy-first, open-source Android app that consolidates multiple AI assistants into a single, customizable interface with built-in ad and tracker blocking.
- Project
deepset-ai/haystack-cookbookHaystack Cookbook is a collection of example notebooks that demonstrate how to use the Haystack framework with different model providers, vector databases, retrieval techniques, and experimental features. It serves as a practical guideline for developers building LLM applications with small, focused demos.