
  <rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
    <channel>
      <title>HANS</title>
      <link>https://www.hansking.cn/blog</link>
      <description>A posts created with Next.js and Tailwind.css</description>
      <language>en-us</language>
      <managingEditor>admin@hansking.cn (HANS)</managingEditor>
      <webMaster>admin@hansking.cn (HANS)</webMaster>
      <lastBuildDate>Fri, 03 Jul 2026 10:00:00 GMT</lastBuildDate>
      <atom:link href="https://www.hansking.cn/tags/评测/feed.xml" rel="self" type="application/rss+xml"/>
      
  <item>
    <guid>https://www.hansking.cn/blog/2026cong-ling-da-jian-LLM-ying-yong-CI-ce-shi-ping-ce-yu-fa-bu-men-jin</guid>
    <title>从零搭建 LLM 应用 CI：测试、评测与发布门禁</title>
    <link>https://www.hansking.cn/blog/2026cong-ling-da-jian-LLM-ying-yong-CI-ce-shi-ping-ce-yu-fa-bu-men-jin</link>
    <description>[{&quot;value&quot;:&quot;从零搭建 LLM 应用 CI：测试、评测与发布门禁&quot;,&quot;url&quot;:&quot;#从零搭建-llm-应用-ci测试评测与发布门禁&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先把测试分成四层&quot;,&quot;url&quot;:&quot;#一先把测试分成四层&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、代码层先测试确定性行为&quot;,&quot;url&quot;:&quot;#二代码层先测试确定性行为&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、链路层用可控替身测试 Agent&quot;,&quot;url&quot;:&quot;#三链路层用可控替身测试-agent&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、黄金数据集是 LLM 应用的回归测试&quot;,&quot;url&quot;:&quot;#四黄金数据集是-llm-应用的回归测试&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、不要用精确字符串比较所有模型答案&quot;,&quot;url&quot;:&quot;#五不要用精确字符串比较所有模型答案&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、把 Prompt 和模型版本当成代码&quot;,&quot;url&quot;:&quot;#六把-prompt-和模型版本当成代码&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、CI 流水线可以这样分段&quot;,&quot;url&quot;:&quot;#七ci-流水线可以这样分段&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、发布门禁要防止回归&quot;,&quot;url&quot;:&quot;#八发布门禁要防止回归&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、让 CI 本身也可靠&quot;,&quot;url&quot;:&quot;#九让-ci-本身也可靠&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;总结：绿色流水线应该代表值得信任&quot;,&quot;url&quot;:&quot;#总结绿色流水线应该代表值得信任&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Fri, 03 Jul 2026 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>CI/CD</category><category>自动化测试</category><category>评测</category>
  </item>

  <item>
    <guid>https://www.hansking.cn/blog/2026tui-li-mo-xing-de-si-wei-lian-ping-ce-yu-jie-guo-jiao-yan</guid>
    <title>推理模型的思维链评测与结果校验</title>
    <link>https://www.hansking.cn/blog/2026tui-li-mo-xing-de-si-wei-lian-ping-ce-yu-jie-guo-jiao-yan</link>
    <description>[{&quot;value&quot;:&quot;推理模型的思维链评测与结果校验&quot;,&quot;url&quot;:&quot;#推理模型的思维链评测与结果校验&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先分清三个概念&quot;,&quot;url&quot;:&quot;#一先分清三个概念&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;最终结果监督：答案对不对&quot;,&quot;url&quot;:&quot;#最终结果监督答案对不对&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;过程监督：中间步骤是否合理&quot;,&quot;url&quot;:&quot;#过程监督中间步骤是否合理&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;可验证中间结果：能否由系统复查&quot;,&quot;url&quot;:&quot;#可验证中间结果能否由系统复查&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;二、不要把“展示思维链”当成评测的终点&quot;,&quot;url&quot;:&quot;#二不要把展示思维链当成评测的终点&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、先设计可评测的任务，而不是先挑模型&quot;,&quot;url&quot;:&quot;#三先设计可评测的任务而不是先挑模型&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、自动校验器是推理评测的核心资产&quot;,&quot;url&quot;:&quot;#四自动校验器是推理评测的核心资产&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;数值与公式&quot;,&quot;url&quot;:&quot;#数值与公式&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;代码&quot;,&quot;url&quot;:&quot;#代码&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;SQL 与数据操作&quot;,&quot;url&quot;:&quot;#sql-与数据操作&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;结构化计划&quot;,&quot;url&quot;:&quot;#结构化计划&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;五、如何评测中间步骤&quot;,&quot;url&quot;:&quot;#五如何评测中间步骤&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、自洽采样有用，但不能迷信多数票&quot;,&quot;url&quot;:&quot;#六自洽采样有用但不能迷信多数票&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、裁判模型只能当辅助裁判&quot;,&quot;url&quot;:&quot;#七裁判模型只能当辅助裁判&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、生产环境要校验结果，而不只是评测结果&quot;,&quot;url&quot;:&quot;#八生产环境要校验结果而不只是评测结果&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、我在评测中最常见的三个误区&quot;,&quot;url&quot;:&quot;#九我在评测中最常见的三个误区&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;十、我的总结：让结论接受事实的审判&quot;,&quot;url&quot;:&quot;#十我的总结让结论接受事实的审判&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Sun, 16 Mar 2025 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>推理模型</category><category>评测</category><category>思维链</category>
  </item>

  <item>
    <guid>https://www.hansking.cn/blog/2026gou-jian-huang-jin-shu-ju-ji-rang-ping-ce-cong-gan-jue-bian-cheng-zheng-ju</guid>
    <title>构建黄金数据集：让评测从感觉变成证据</title>
    <link>https://www.hansking.cn/blog/2026gou-jian-huang-jin-shu-ju-ji-rang-ping-ce-cong-gan-jue-bian-cheng-zheng-ju</link>
    <description>[{&quot;value&quot;:&quot;构建黄金数据集：让评测从感觉变成证据&quot;,&quot;url&quot;:&quot;#构建黄金数据集让评测从感觉变成证据&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先定义数据集要回答什么问题&quot;,&quot;url&quot;:&quot;#一先定义数据集要回答什么问题&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、样本要覆盖真实分布和危险边界&quot;,&quot;url&quot;:&quot;#二样本要覆盖真实分布和危险边界&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;三、期望答案要写成可判断的标准&quot;,&quot;url&quot;:&quot;#三期望答案要写成可判断的标准&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;四、标注需要规则，也需要人&quot;,&quot;url&quot;:&quot;#四标注需要规则也需要人&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;五、时间切分和客户切分防止污染&quot;,&quot;url&quot;:&quot;#五时间切分和客户切分防止污染&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;六、版本化和变更记录不能省&quot;,&quot;url&quot;:&quot;#六版本化和变更记录不能省&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;七、指标要和数据集结构对应&quot;,&quot;url&quot;:&quot;#七指标要和数据集结构对应&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;八、黄金集要服务模型、RAG 和 Agent&quot;,&quot;url&quot;:&quot;#八黄金集要服务模型rag-和-agent&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;九、数据集不是建完就结束&quot;,&quot;url&quot;:&quot;#九数据集不是建完就结束&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;总结：黄金数据集是一种工程记忆&quot;,&quot;url&quot;:&quot;#总结黄金数据集是一种工程记忆&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Mon, 09 Feb 2026 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>数据集</category><category>评测</category><category>工程实践</category>
  </item>

  <item>
    <guid>https://www.hansking.cn/blog/2026sheng-chan-ji-RAG-li-xian-ping-ce-zai-xian-jian-kong-yu-zeng-liang-suo-yin</guid>
    <title>生产级 RAG：离线评测、在线监控与增量索引</title>
    <link>https://www.hansking.cn/blog/2026sheng-chan-ji-RAG-li-xian-ping-ce-zai-xian-jian-kong-yu-zeng-liang-suo-yin</link>
    <description>[{&quot;value&quot;:&quot;生产级 RAG：离线评测、在线监控与增量索引&quot;,&quot;url&quot;:&quot;#生产级-rag离线评测在线监控与增量索引&quot;,&quot;depth&quot;:1},{&quot;value&quot;:&quot;一、先把生产级 RAG 想清楚&quot;,&quot;url&quot;:&quot;#一先把生产级-rag-想清楚&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;二、离线评测：上线之前先知道它好不好&quot;,&quot;url&quot;:&quot;#二离线评测上线之前先知道它好不好&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;1. 先做黄金数据集，而不是先选评测框架&quot;,&quot;url&quot;:&quot;#1-先做黄金数据集而不是先选评测框架&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;2. 把检索和生成分开打分&quot;,&quot;url&quot;:&quot;#2-把检索和生成分开打分&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;3. 评测必须能比较版本&quot;,&quot;url&quot;:&quot;#3-评测必须能比较版本&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;三、在线监控：让系统自己告诉你哪里坏了&quot;,&quot;url&quot;:&quot;#三在线监控让系统自己告诉你哪里坏了&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;1. 先监控基础设施，再监控答案质量&quot;,&quot;url&quot;:&quot;#1-先监控基础设施再监控答案质量&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;四、增量索引：知识库更新不能每次从头再来&quot;,&quot;url&quot;:&quot;#四增量索引知识库更新不能每次从头再来&quot;,&quot;depth&quot;:2},{&quot;value&quot;:&quot;更新过程要可恢复&quot;,&quot;url&quot;:&quot;#更新过程要可恢复&quot;,&quot;depth&quot;:3},{&quot;value&quot;:&quot;五、我最后留下的一份上线清单&quot;,&quot;url&quot;:&quot;#五我最后留下的一份上线清单&quot;,&quot;depth&quot;:2}]</description>
    <pubDate>Fri, 27 Feb 2026 10:00:00 GMT</pubDate>
    <author>admin@hansking.cn (HANS)</author>
    <category>AI</category><category>LLM</category><category>RAG</category><category>评测</category><category>工程实践</category>
  </item>

    </channel>
  </rss>
