大模型&智能体分享大纲
<h3><strong>课程总览</strong></h3><p>· <strong><span style="font-family: "Songti SC Regular"">时长</span></strong><span style="font-family: "Songti SC Regular"">：</span><span style="font-family: "Songti SC Regular"">40小时（含</span><span style="font-family: "Songti SC Regular"">10</span><span style="font-family: "Songti SC Regular"">小时实践项目）</span><br><strong><span style="font-family: "Songti SC Regular"">目标</span></strong><span style="font-family: "Songti SC Regular"">：掌握大模型全生命周期开发能力，具备独立优化和部署行业模型的能力</span></p>
<h3><strong>大模型篇</strong></h3>
<h3><strong>模块</strong><strong>1</strong><strong>：模型</strong><strong>原理</strong><strong>（</strong><strong>5</strong><strong>小时）</strong></h3>
<p class="p"><strong>目标</strong>：掌握模型的基础原理更好的理解内容</p>
<p>1. <strong>Attention</strong><strong><span style="font-family: "Songti SC Regular"">讲解</span></strong><strong><span style="font-family: "Songti SC Regular"">（</span></strong><strong><span style="font-family: "Songti SC Regular"">1h</span></strong><strong><span style="font-family: "Songti SC Regular"">）</span></strong></p>
<p>2. <strong><span style="font-family: "Songti SC Regular"">Self—</span></strong><strong>Attention</strong><strong><span style="font-family: "Songti SC Regular"">代码实现讲解</span></strong><strong><span style="font-family: "Songti SC Regular"">（</span></strong><strong><span style="font-family: "Songti SC Regular"">1</span></strong><strong><span style="font-family: "Songti SC Regular"">h）</span></strong></p>
<p>3. <strong>Transformer</strong><strong><span style="font-family: "Songti SC Regular"">结构讲解（</span><span style="font-family: "Songti SC Regular"">1h）</span></strong></p>
<p>4. <strong>Transformer</strong><strong><span style="font-family: "Songti SC Regular"">模块代码实现讲解（</span><span style="font-family: "Songti SC Regular"">1h）</span></strong></p>
<p>5. <strong><span style="font-family: "Songti SC Regular"">GPT、T5、deepseek大模型差异化对比（1h）</span></strong></p>
<h3><strong>模块</strong><strong>2</strong><strong>：模型微调</strong><strong>与部署</strong><strong>（</strong><strong>5</strong><strong>小时）</strong></h3>
<p class="p"><strong>目标</strong>：掌握不同场景下的模型优化策略</p>
<p>1.微调（2h）</p>
<p>微调是什么</p>
<p>为什么要微调</p>
<p>如何进行微调</p>
<p>2.Adapter（2h）</p>
<p> Adapter是什么</p>
<p> Adapter优势和不足</p>
<p> Adapter 微调实例</p>
<p>3.Prompt-Tuning（2h）</p>
<p>Prompt-Tuning是什么</p>
<p>P-Tuning</p>
<p>Prefix-Tuning</p>
<p> 4.P-Tuning（4h）</p>
<p>P-Tuning是什么</p>
<p>P-Tuning的优势</p>
<p>P-Tuning的实现过程 </p>
<p>LoRA原理以及案例</p>
<p>5.qwen模型微调及部署案例（1h）</p>
<p>qwen全参数微调 </p>
<p>qwen的LoRA微调</p>
<p>6.deepseek模型微调及部署案例（1h）</p>
<p>deepseek全参数微调 </p>
<p>deepseek的LoRA微调</p>
<h3><strong>模块</strong><strong>3</strong><strong>：模型评估与优化（</strong><strong>5</strong><strong>小时）</strong></h3>
<p class="p"><strong>目标</strong>：建立科学的模型评价体系</p>
<p><span style="font-family: "Songti SC Regular"">1.</span><span style="font-family: "Songti SC Regular"">评估指标体系（</span><span style="font-family: "Songti SC Regular"">2</span><span style="font-family: "Songti SC Regular"">h）</span></p>
<p><span style="font-family: "Songti SC Regular"">文本：</span><span style="font-family: "Songti SC Regular"">BLEU/ROUGE/BERTScore</span></p>
<p><span style="font-family: "Songti SC Regular"">对话：</span><span style="font-family: "Songti SC Regular"">USR/DSSM</span></p>
<p>2.评估内容具体方面（3h）</p>
<p>内容质量</p>
<p>内容安全性</p>
<p>内容可控性</p>
<p>内容个性化</p>
<p>业务价值评估</p>
<p>3.<span style="font-family: "Songti SC Regular"">典型评估案例</span></p>
<p><span style="font-family: "Songti SC Regular"">医疗问诊模型</span></p>
<p><span style="font-family: "Songti SC Regular"">核心指标：诊断准确率（</span><span style="font-family: "Songti SC Regular"">vs 主治医师）、用药建议合规率</span></p>
<p><span style="font-family: "Songti SC Regular"">特殊测试：罕见病案例处理能力、紧急情况响应速度</span></p>
<p><span style="font-family: "Songti SC Regular"">金融报告生成</span></p>
<p><span style="font-family: "Songti SC Regular"">核心指标：数据一致性（</span><span style="font-family: "Songti SC Regular"">vs 原始财报）、关键指标覆盖率</span></p>
<p><span style="font-family: "Songti SC Regular"">特殊测试：市场敏感信息脱敏、多语言财报生成质量</span></p>
<p><span style="font-family: "Songti SC Regular"">教育辅导系统</span></p>
<p><span style="font-family: "Songti SC Regular"">核心指标：知识点覆盖度、学生留存率</span></p>
<p><span style="font-family: "Songti SC Regular"">特殊测试：错题解析逻辑性、个性化学习路径推荐合理性</span></p>
<h3><strong>模块</strong><strong>4</strong><strong>：行业</strong><strong><span style="font-family: "Songti SC Regular"">案例解析（</span><span style="font-family: "Songti SC Regular"">5h）</span></strong></h3>
<h6>1. AIGC在搜索广告创意场景中的应用</h6>
<h6>创意大模型生成标题</h6>
<h6>基于原生化营销知识强化的开放式广告标题生成</h6>
<h6><span style="font-family: "Songti SC Regular"">基于</span><span style="font-family: "Songti SC Regular"">LoRA的医疗行业生成式标题优化</span></h6>
<p>2. 基于关键词文本生成大模型的广告检索</p>
<p>基于关键词文本生成的单任务建模</p>
<p>基于关键词生成多任务统一建模</p>
<p>基于百亿模型的关键词多结果生成任务</p>
<p>关键词受限生成和生成判别一体化</p>
<p> <span style="font-family: "Songti SC Regular"">基于生成式模型的短语匹配任务建模</span></p>
<p>3.生成判别一体化大模型广告召回</p>
<p>字典树限定生成大模型</p>
<p>生成判别一体化模型</p>
<p>基于生成判别模型的短语匹配任务建模</p>
<p>4.基于生成式的端到端定向广告召回</p>
<p>基于广告主生成式检索</p>
<p>基于<span style="font-family: "Songti SC Regular"">语义</span><span style="font-family: "Songti SC Regular"">ID生成式检索</span></p>
<p>基于业务点生成式大模型</p>
<p> </p>
<p><strong>RAG（5h）</strong></p>
<p>1. <strong>背景介绍</strong></p>
<p>1.1. <strong>企业知识库特点</strong></p>
<p>1.2. <strong>RAG 技术概述</strong></p>
<p><strong>1.3.RAG 的领域应用</strong></p>
<p><strong>2.RAG 整体优化方案介绍</strong></p>
<p><strong>2.1. 构建索引优化</strong></p>
<p><strong>2.2. 检索优化</strong></p>
<p><strong>2.3. 向量与排序模型微调</strong></p>
<p><strong>3.RAG 技术</strong></p>
<p><strong>3.1. 嵌入技术 Embedding 模型</strong></p>
<p><strong>3.2. 向量数据库</strong></p>
<p><strong>3.3. 业务数据向量化</strong></p>
<p><strong>4.1RAG在办公领域的企业实践</strong></p>
<p><strong>4.2图RAG在企业知识服务中的落地</strong></p>
<p><strong> </strong></p>
<p><strong>Agent（15h）</strong></p>
<p><span style="font-family: "Hiragino Sans GB"">大模型</span> Agent</p>
<p>Concepts 概念</p>
<p>OpenAI 函数智能体</p>
<p>OpenAI 函数调用</p>
<p>XML 智能体</p>
<p>JSON 结构智能体</p>
<p>ReAct 智能体</p>
<p><strong><span style="font-family: "Songti SC Bold"">多自能体协同</span></strong></p>
<p><strong><span style="font-family: "Songti SC Bold"">记忆</span></strong></p>
<p><strong> </strong></p>
</div>
<div id="MySignature" role="contentinfo">
<p>本文来自博客园，作者：limingqi，转载请注明原文链接：https://www.cnblogs.com/limingqi/p/19480796</p><br><br>
来源：https://www.cnblogs.com/limingqi/p/19480796
頁:
[1]