| 课表 | 描述 | 课程资料 | 任务 |
|---|---|---|---|
| 第1节 | 大模型基础：理论与技术的演进 - 初探大模型：起源与发展 - 预热篇：解码注意力机制 - 变革里程碑：Transformer的崛起 - 走向不同：GPT与BERT的选择 |
建议阅读： - Attention Mechanism: Neural Machine Translation by Jointly Learning to Align and Translate - An Attentive Survey of Attention Models - Transformer：Attention is All you Need - [BERT：Pre-training of Deep Bidirectional Transformers for Language Understanding(https://arxiv.org/abs/1810.04805) |
[作业] |
| 第2节 | GPT 模型家族：从始至今 - 从GPT-1到GPT-3.5：一路的风云变幻 - ChatGPT：赢在哪里 - GPT-4：一个新的开始 提示学习（Prompt Learning） - 思维链（Chain-of-Thought, CoT）：开山之作 - 自洽性（Self-Consistency）：多路径推理 - 思维树（Tree-of-Thoughts, ToT）：续写佳话 |
建议阅读： - GPT-1: Improving Language Understanding by Generative Pre-training - GPT-2: Language Models are Unsupervised Multitask Learners - GPT-3: Language Models are Few-Shot Learners 额外阅读： - GPT-4: Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE - GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models - Sparks of Artificial General Intelligence: Early experiments with GPT-4 |
[作业] |
| 第3节 | 大模型开发基础：OpenAI Embedding - 通用人工智能的前夜 - "三个世界"和"图灵测试" - 计算机数据表示 - 表示学习和嵌入 Embeddings Dev 101 - 课程项目：GitHub openai-quickstart - 快速上手 OpenAI Embeddings |
建议阅读： - Representation Learning: A Review and New Perspectives - Word2Vec: Efficient Estimation of Word Representations in Vector Space - GloVe: Global Vectors for Word Representation 额外阅读： - Improving Distributional Similarity with Lessons Learned from Word Embeddings - Evaluation methods for unsupervised word embeddings |
[作业] 代码： [embedding] |
| 第4节 | OpenAI 大模型开发与应用实践 - OpenAI大型模型开发指南 - OpenAI 语言模型总览 - OpenAI GPT-4, GPT-3.5, GPT-3, Moderation - OpenAI Token 计费与计算 OpenAI API 入门与实战 - OpenAI Models API - OpenAI Completions API - OpenAI Chat Completions API - Completions vs Chat Completions OpenAI 大模型应用实践 - 文本内容补全初探（Text Completion） - 聊天机器人初探（Chat Completion） |
建议阅读： - OpenAI Models - OpenAI Completions API - OpenAI Chat Completions API |
代码： [models] [tiktoken] |
| 第5节 | AI大模型应用最佳实践 - 如何提升GPT模型使用效率与质量 - AI大模型应用最佳实践 - 文本创作与生成 - 文章摘要和总结 - 小说生成与内容监管 - 分步骤执行复杂任务 - 评估模型输出质量 - 构造训练标注数据 - 代码调试助手 - 新特性： Function Calling 介绍与实战 |
建议阅读 - GPT Best Practices - Function Calling |
代码： Function Calling |
| 第6节 | 实战：OpenAI-Translator - OpenAI-Translator 市场需求分析 - OpenAI-Translator 产品定义与功能规划 - OpenAI-Translator 技术方案与架构设计 - OpenAI 模块设计 - OpenAI-Translator 实战 |
代码： pdfplumber |
|
| 第7节 | 实战：ChatGPT Plugin 开发 - ChatGPT Plugin 开发指南 - ChatGPT Plugin 介绍 - ChatGPT Plugin 介绍 - 样例项目：待办（Todo）管理插件 - 实战样例部署与测试 - ChatGPT 开发者模式 - 实战：天气预报（Weather Forecast）插件开发 - Weather Forecast Plugin 设计与定义 - 天气预报函数服务化 - 第三方天气查询平台对接 - 实战 Weather Forecast Plugin - Function Calling vs ChatGPT plugin |
代码： [todo list] [Weather Forecast] |
|
| 第8节 | 大模型应用开发框架 LangChain (上) - LangChain 101 - LangChain 是什么 - 为什么需要 LangChain - LangChain 典型使用场景 - LangChain 基础概念与模块化设计 - LangChain 核心模块入门与实战 - 标准化的大模型抽象：Mode I/O - 模板化输入：Prompts - 语言模型：Models - 规范化输出：Output Parsers |
代码： [model io] |
|
| 第9节 | 大模型应用开发框架 LangChain (中) - 大模型应用的最佳实践 Chains - 上手你的第一个链：LLM Chain - 串联式编排调用链：Sequential Chain - 处理超长文本的转换链：Transform Chain - 实现条件判断的路由链：Router Chain - 赋予应用记忆的能力： Memory - Momory System 与 Chain 的关系 - 记忆基类 BaseMemory 与 BaseChatMessageMemory - 服务聊天对话的记忆系统 - ConversationBufferMemory - ConversationBufferWindowMemory - ConversationSummaryBufferMemory |
代码： [chains] [memory] |
|
| 第10节 | 大模型应用开发框架 LangChain (下) - 框架原生的数据处理流 Data Connection - 文档加载器（Document Loaders） - 文档转换器（Document Transformers） - 文本向量模型（Text Embedding Models） - 向量数据库（Vector Stores） - 检索器（Retrievers） - 构建复杂应用的代理系统 Agents - Agent 理论基础：ReAct - LLM 推理能力：CoT, ToT - LLM 操作能力：WebGPT, SayCan - LangChain Agents 模块设计与原理剖析 - Module: Agent, Tools, Toolkits, - Runtime: AgentExecutor, PlanAndExecute , AutoGPT, - 上手第一个Agent：Google Search + LLM - 实战 ReAct：SerpAPI + LLM-MATH |
代码： [data connection] [agents] |
|
| 第11节 | 实战： LangChain 版 OpenAI-Translator v2.0 - 深入理解 Chat Model 和 Chat Prompt Template - 温故：LangChain Chat Model 使用方法和流程 - 使用 Chat Prompt Template 设计翻译提示模板 - 使用 Chat Model 实现双语翻译 - 使用 LLMChain 简化构造 Chat Prompt - 基于 LangChain 优化 OpenAI-Translator 架构设计 - 由 LangChain 框架接手大模型管理 - 聚焦应用自身的 Prompt 设计 - 使用 TranslationChain 实现翻译接口 - 更简洁统一的配置管理 - OpenAI-Translator v2.0 功能特性研发 - 基于Gradio的图形化界面设计与实现 - 基于 Flask 的 Web Server 设计与实现 |
代码： [openai-translator] |
|
| 第12节 | 实战： LangChain 版Auto-GPT - Auto-GPT 项目定位与价值解读 - Auto-GPT 开源项目介绍 - Auto-GPT 定位：一个自主的 GPT-4 实验 - Auto-GPT 价值：一种基于 Agent 的 AGI 尝试 - LangChain 版 Auto-GPT 技术方案与架构设计 - 深入理解 LangChain Agents - LangChain Experimental 模块 - Auto-GPT 自主智能体设计 - Auto-GPT Prompt 设计 - Auto-GPT Memory 设计 - 深入理解 LangChain VectorStore - Auto-GPT OutputParser 设计 - 实战 LangChain 版 Auto-GPT |
代码： [autogpt] |
|
| 第13节 | Sales-Consultant 业务流程与价值分析 - Sales-Consultant 技术方案与架构设计 - 使用 GPT-4 生成销售话术 - 使用 FAISS 向量数据库存储销售问答话术 - 使用 RetrievalQA 检索销售话术数据 - 使用 Gradio 实现聊天机器人的图形化界面 - 实战 LangChain 版 Sales-Consultant |
代码： [sales_chatbot] |
|
| 第14节 | 大模型时代的开源与数据协议 - 什么是开源？ - 广泛使用的开源协议和数据协议 - Llama 是不是伪开源？ - ChatGLM2-6B 的开源协议 大语言模型的可解释性 - 提高模型决策过程的透明度 - Stanford Alpaca 的相关研究 大语言模型应用的法规合规性 - 中国大陆：生成式人工智能服务备案 - 国际化：数据隐私与保护（以 GDPR 为例） - 企业合规性应对要点 |
||
| 第15节 | 大模型时代的Github：Hugging Face - Hugging Face 是什么？ - Hugging Face Transformers 库 - Hugging Face 开源社区：Models, Datasets, Spaces, Docs - 大模型横向对比 - Open LLM Leaderboard（大模型天梯榜） 显卡选型推荐指南 - GPU vs 显卡 - GPU Core vs AMD CU - CUDA Core vs Tensor Core - N卡的架构变迁 - 显卡性能天梯榜 |
||
| 第16节 | 清华 GLM 大模型家族 - 最强基座模型 GLM-130B - 增强对话能力 ChatGLM - 开源聊天模型 ChatGLM2-6B - 联网检索能力 WebGLM - 初探多模态 VisualGLM-6B - 代码生成模型 CodeGeex2 ChatGLM2-6B 大模型应用开发 - ChatGLM2-6B 私有化部署 - HF Transformers Tokenizer - HF Transformers Model - 将模型同步至 Hugging Face - 使用 Gradio 赋能 ChatGLM2-6B 图形化界面 < |
| Lesson | Description | Course Materials | Events |
|---|---|---|---|
| Lesson 1 | Fundamentals of Large Models: Evolution of Theory and Technology - An Initial Exploration of Large Models: Origin and Development - Warm-up: Decoding Attention Mechanism - Milestone of Transformation: The Rise of Transformer - Taking Different Paths: The Choices of GPT and Bert |
Suggested Readings: - Attention Mechanism: Neural Machine Translation by Jointly Learning to Align and Translate - An Attentive Survey of Attention Models - Transformer: Attention is All you Need - BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding |
[Homework] |
| Lesson 2 | The GPT Model Family: From Start to Present - From GPT-1 to GPT-3.5: The Evolution - ChatGPT: Where It Wins - GPT-4: A New Beginning Prompt Learning - Chain-of-Thought (CoT): The Pioneering Work - Self-Consistency: Multi-path Reasoning - Tree-of-Thoughts (ToT): Continuing the Story |
Suggested Readings: - GPT-1: Improving Language Understanding by Generative Pre-training - GPT-2: Language Models are Unsupervised Multitask Learners - GPT-3: Language Models are Few-Shot Learners Additional Readings: - GPT-4: Architecture, Infrastructure, Training Dataset, Costs, Vision, MoE - GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models - Sparks of Artificial General Intelligence: Early experiments with GPT-4 |
[Homework] |
| Lesson 3 | Fundamentals of Large Model Development: OpenAI Embedding - The Eve of General Artificial Intelligence - "Three Worlds" and "Turing Test" - Computer Data Representation - Representation Learning and Embedding Embeddings Dev 101 - Course Project: GitHub openai-quickstart - Getting Started with OpenAI Embeddings |
Suggested Readings: - Representation Learning: A Review and New Perspectives - Word2Vec: Efficient Estimation of Word Representations in Vector Space - GloVe: Global Vectors for Word Representation Additional Readings: - Improving Distributional Similarity with Lessons Learned from Word Embeddings - Evaluation methods for unsupervised word embeddings |
[Homework] Code: [embedding] |
| Lesson 4 | OpenAI Large Model Development and Application Practice - OpenAI Large Model Development Guide - Overview of OpenAI Language Models - OpenAI GPT-4, GPT-3.5, GPT-3, Moderation - OpenAI Token Billing and Calculation OpenAI API Introduction and Practice - OpenAI Models API - OpenAI Completions API - OpenAI Chat Completions API - Completions vs Chat Completions OpenAI Large Model Application Practice - Initial Exploration of Text Completion - Initial Exploration of Chatbots |
Suggested Readings: - OpenAI Models - OpenAI Completions API - OpenAI Chat Completions API |
Code: [models] [tiktoken] |
| Lesson 5 | Best Practices for Applying Large AI Models - How to Improve the Efficiency and Quality of GPT Model Use - Best Practices for Applying Large AI Models - Text Creation and Generation - Article Abstract and Summary - Novel Generation and Content Supervision - Executing Complex Tasks Step by Step - Evaluating the Quality of Model Output - Constructing Training Annotation Data - Code Debugging Assistant - New Features： Function Calling Introduction and Practical Application |
Suggested Readings - GPT Best Practices - Function Calling |
Code： Function Calling |
| Lesson 6 | Practical: OpenAI-Translator - Market demand analysis for OpenAI-Translator - Product definition and feature planning for OpenAI-Translator - Technical solutions and architecture design for OpenAI-Translator - OpenAI module design - OpenAI-Translator practical application |
Code: pdfplumber |
|
| Lesson 7 | ChatGPT Plugin Development Guide - Introduction to ChatGPT Plugin - Sample project: Todo management plugin - Deployment and testing of practical examples - ChatGPT developer mode - Practical: Weather Forecast plugin development - Weather Forecast Plugin design and definition - Weather Forecast function service - Integration with third-party weather query platform - Practical Weather Forecast Plugin |
Code: [todo list] [weather forecast] |
|
| Lesson 8 | LLM Application Development Framework LangChain (Part 1) - LangChain 101 - What is LangChain - Why LangChain is Needed - Typical Use Cases of LangChain - Basic Concepts and Modular Design of LangChain - Introduction and Practice of LangChain Core Modules - Standardized Large-Scale Model Abstraction: Mode I/O - Template Input: Prompts - Language Model: Models - Standardized Output: Output Parsers |
Code: [model io] |
|
| Lesson 9 | LLM Application Development Framework LangChain (Part 2) - Best Practices for LLM Chains - Getting Started with Your First Chain: LLM Chain - Sequential Chain: A Chained Call with Sequential Arrangement - Transform Chain: A Chain for Processing Long Texts - Router Chain: A Chain for Implementing Conditional Judgments - Memory: Endowing Applications with Memory Capabilities - The Relationship between Memory System and Chain - BaseMemory and BaseChatMessageMemory: Memory Base Classes - Memory System for Service Chatting - ConversationBufferMemory - ConversationBufferWindowMemory - ConversationSummaryBufferMemory |
Code: [chains] [memory] |
|
| Lesson 10 | LLM Application Development Framework LangChain (Part 3) - Native data processing flow of the framework: Data Connection - Document Loaders - Document Transformers - Text Embedding Models - Vector Stores - Retrievers - Agent Systems for Building Complex Applications: Agents - Theoretical Foundation of Agents: ReAct - LLM Reasoning Capabilities: CoT, ToT - LLM Operation Capabilities: WebGPT, SayCan - LangChain Agents Module Design and Principle Analysis - Module: Agent, Tools, Toolkits - Runtime: AgentExecutor, PlanAndExecute, AutoGPT - Getting Started with Your First Agent: Google Search + LLM - Practice with ReAct: SerpAPI + LLM-MATH |
Code: [data connection] [agents] |
|
| Lesson 11 | Practical: LangChain version OpenAI-Translator v2.0 - In-depth understanding of Chat Model and Chat Prompt Template - Review: LangChain Chat Model usage and process - Design translation prompt templates using Chat Prompt Template - Implement bilingual translation using Chat Model - Simplify Chat Prompt construction using LLMChain - Optimize OpenAI-Translator architecture design based on LangChain - Hand over large model management to LangChain framework - Focus on application-specific Prompt design - Implement translation interface using TranslationChain - More concise and unified configuration management - Development of OpenAI-Translator v2.0 feature - Design and implementation of graphical interface based on Gradio - Design and implementation of Web Server based on Flask |
Code: [openai-translator] |
|
| Lesson 12 | Practical: LangChain version Auto-GPT - Auto-GPT project positioning and value interpretation - Introduction to Auto-GPT open source project - Auto-GPT positioning: an independent GPT-4 experiment - Auto-GPT value: an attempt at AGI based on Agent - LangChain version Auto-GPT technical solution and architecture design - In-depth understanding of LangChain Agents - LangChain Experimental module - Auto-GPT autonomous agent design - Auto-GPT Prompt design - Auto-GPT Memory design - In-depth understanding of LangChain VectorStore - Auto-GPT OutputParser design - Practical LangChain version Auto-GPT |
Code: [autogpt] |
|
| Lesson 13 | Sales-Consultant business process and value analysis - Technical solution and architecture design of Sales-Consultant - Use GPT-4 to generate sales pitches - Store sales Q&A pitches in FAISS vector database - Retrieve sales pitches data using RetrievalQA - Implement chatbot graphical interface using Gradio - Practical LangChain version Sales-Consultant |
Code： [sales_chatbot] |
|
| Lesson 14 | Era of large models: Open source and data protocols - What is open source? - Widely used open source and data protocols - Is Llama pseudo-open source? - Open source protocol of ChatGLM2-6B Interpretability of large language models - Enhancing transparency in model decision-making - Related research of Stanford Alpaca Regulatory compliance of large language model applications - Mainland China: Registration of generative AI services - International: Data privacy and protection (taking GDPR as an example) - Key points of corporate compliance |
||
| Lesson 15 | Github in the era of large models: Hugging Face - What is Hugging Face? - Hugging Face Transformers library - Hugging Face open community: Models, Datasets, Spaces, Docs - Comparative analysis of large models - Open LLM Leaderboard (Large Model Ladder) Graphics card selection guide - GPU vs Graphics card - GPU Core vs AMD CU - CUDA Core vs Tensor Core - Evolution of Nvidia architectures - Graphics card performance ladder |
||
| Lesson 16 | Tsinghua GLM large model family - Strongest base model GLM-130B - Enhanced dialogue capability ChatGLM - Open source chat model ChatGLM2-6B - Internet search capability WebGLM - Initial exploration of multimodal VisualGLM-6B - Code generation model CodeGeex2 Application development of ChatGLM2-6B large model - Private deployment of ChatGLM2-6B - HF Transformers Tokenizer - HF Transformers Model - Synchronize the model to Hugging Face - Empower ChatGLM2-6B graphical interface using Gradio - Fine-tuning of ChatGLM2-6B model - Practical assignment: Implement graphical interface of openai-translator based on ChatGLM2-6B |