PART 4安装使用十分钟跑起来环境要求Python 3.12 或 3.13 uvtorch 尚无 3.14 wheel。GPU 可选CUDA / ROCm / Apple SiliconMLX 自动启用。第一步浏览器先试可选官方 HuggingFace Space 跑着 Kev-4B 和 Kev-0.8B零安装huggingface.co/spaces/jaredpalmer/kev。第二步本地起服务gitclone https://github.com/jaredpalmer/kev.gitcdkev uvsync--extraserve uv run--extraserve python-mkev.serve--runjaredpalmer/kev-4b--port8009首次运行自动下载适配器与底座模型。--run也接受本地 checkpoint 目录或 Hub revision。第三步发一个请求curl-slocalhost:8009/v1/systemone-Hcontent-type: application/json-d{ state: Shoes arrived two weeks late and in the wrong size. Also I see two charges on my card., model: kev-latest, questions: { department: {type: choice, instructions: Which team should handle this?, criteria: {returns: Exchanges, refunds, wrong or damaged items, shipping: Delivery status, delays, lost packages, billing: Charges, invoices, payment problems}}, escalate: {type: noul, instructions: Does this need urgent human attention?}, frustration: {type: score, instructions: How frustrated is the customer?, criteria: [Calm, Frustrated, Very angry]} }}响应Kev-4BApple M5 bf16{answers:{department:{type:choice,choice:returns,confidence:0.21,probabilities:{returns:0.47,shipping:0.28,billing:0.25}},escalate:{type:noul,noul:0.93},frustration:{type:score,score:1.44,confidence:0.34,probabilities:{0:0.00,1:0.56,2:0.44}}},usage:{input_tokens:101,output_tokens:161},latency_ms:495}注意 department 的概率分布如实反映了工单里的三个诉求——返回概率而不是单一标签就是为了让你的代码按置信度路由。第四步Python 接入已在用 Jev 的零改造fromtypesafe_sdkimportChoice,Noul,Score,TypeSafeClient clientTypeSafeClient(api_keylocal,base_urlhttp://127.0.0.1:8009,modelkev-latest,)responseclient.system_one(stateI was charged twice. Please fix this ASAP.,questions{billing:Noul(instructionsIs this ticket about billing?),tone:Choice(instructionsWhat is the customers tone?,criteria{calm:None,frustrated:None,angry:None}),urgency:Score(instructionsHow urgent is this ticket?,criteria[can wait,this week,today]),},)云端部署Modal一条命令pipinstallmodalmodal setupcurl-LOhttps://raw.githubusercontent.com/jaredpalmer/kev/main/skills/kev-deploy/scripts/kev_serve.pyKEV_API_KEY$(openssl rand-hex24)modal deploy kev_serve.py得到https://workspace--kev-api.modal.runBearer 鉴权空闲缩容到零不计费冷启动约 35 秒。常用环境变量变量作用KEV_TEMPERATURE1.0关闭校准返回原始概率KEV_DTYPEfp32用与官方评测完全一致的 fp32 路径KEV_API_KEY要求 Bearer 鉴权KEV_DATE_FACTS1给 state 里的日期对自动附加天数差补日期减法短板我实测的环境坑Windows 用户必读代理环境半开挂死本机开了 HTTP 代理时训练/服务子进程会对 HuggingFace 做 HEAD 请求然后无限挂起CPU/GPU 全 0 空转。解法模型权重预先本地化训练与服务一律HF_HUB_OFFLINE1并清空代理变量。POSIX-only 导入train.py/serve.py顶层import resource测试引fcntl。Windows 下用 runpy 预注入 no-op stub这两个模块在上游只用于峰值内存统计和文件锁stub 不影响训练数值不要改上游仓库。HuggingFace 直连不通国内环境设HF_ENDPOINThttps://hf-mirror.com。--max_steps 0的语义是「无上限」而不是「不训练」。其他入口Playgroundcd playground npm install npm run dev网页里改 state 和问题、对比「打包问 vs 分开问」、做选项顺序扰动还有个拿棋盘当输入的国际象棋 demo。Coding Agent 一键微调npx skills add jaredpalmer/kevkev-finetuneagent 会面谈、找问题、造数据、在 Modal 上训完并部署端点H100 一次约 1 美元。