设计大语言模型应用（影印版）
设计大语言模型应用（影印版）
Suhas Pai 著
出版时间：2025年07月
页数：343
“你翻阅无数论文，尝试各种工具和基准测试，花费数周甚至数月时间才弄清楚LLM的运作机制，而这一切全都涵盖在了这本书中。强烈推荐。”
——Megan Risdal
Kaggle（Google）首席产品经理
“本书是构建先进AI系统的一堂大师课。”
——Jay Alammar
Hands-On Large Language Models合著者
“这本精心策划的佳作涵盖了该领域所有重要的概念与实用知识，真是难得一见。”
——Madhav Singhal
AutoComputer首席执行官

大型语言模型（LLM）已被证明是解决各种任务的强大工具，也得到了企业界的关注。但从演示和原型过渡到真正可用的成熟应用却并不容易。本书旨在弥合这一差距，为实践者提供构建语言模型驱动产品所需的工具、技术和实操指南。
经验丰富的机器学习研究员Suhas Pai就如何利用LLM解决实际应用场景中的问题，以及如何应对常见的失败模式，提供了实用建议。你将深入了解构成语言模型的各项要素，探索诸如微调等多种定制化技术，学习RAG（检索增强生成）和智能体等应用范式。
● 理解如何为训练和微调准备数据集
● 建立对Transformer架构及其变体的直觉认识
● 将预训练语言模型适配于你自己的领域和使用场景
● 掌握微调、领域自适应与推理优化的有效技术
● 使语言模型能够与外部工具和数据对接，并将其集成进现有的软件生态系统
  1. Preface
  2. Part I. LLM Ingredients
  3. 1. Introduction
  4. Defining LLMs
  5. A Brief History of LLMs
  6. The Impact of LLMs
  7. LLM Usage in the Enterprise
  8. Prompting
  9. Accessing LLMs Through an API
  10. Strengths and Limitations of LLMs
  11. Building Your First Chatbot Prototype
  12. From Prototype to Production
  13. Summary
  14. 2. Pre-Training Data
  15. Ingredients of an LLM
  16. Pre-Training Data Requirements
  17. Popular Pre-Training Datasets
  18. Synthetic Pre-Training Data
  19. Training Data Preprocessing
  20. Effect of Pre-Training Data on Downstream Tasks
  21. Bias and Fairness Issues in Pre-Training Datasets
  22. Summary
  23. 3. Vocabulary and Tokenization
  24. Vocabulary
  25. Tokenizers
  26. Tokenization Pipeline
  27. Summary
  28. 4. Architectures and Learning Objectives
  29. Preliminaries
  30. Representing Meaning
  31. The Transformer Architecture
  32. Loss Functions
  33. Intrinsic Model Evaluation
  34. Transformer Backbones
  35. Learning Objectives
  36. Pre-Training Models
  37. Summary
  38. Part II. Utilizing LLMs
  39. 5. Adapting LLMs to Your Use Case
  40. Navigating the LLM Landscape
  41. How to Choose an LLM for Your Task
  42. Loading LLMs
  43. Decoding Strategies
  44. Running Inference on LLMs
  45. Structured Outputs
  46. Model Debugging and Interpretability
  47. Summary
  48. 6. Fine-Tuning
  49. The Need for Fine-Tuning
  50. Fine-Tuning: A Full Example
  51. Fine-Tuning Datasets
  52. Summary
  53. 7. Advanced Fine-Tuning Techniques
  54. Continual Pre-Training
  55. Parameter-Efficient Fine-Tuning
  56. Combining Multiple Models
  57. Summary
  58. 8. Alignment Training and Reasoning
  59. Defining Alignment Training
  60. Reinforcement Learning
  61. Hallucinations
  62. Mitigating Hallucinations
  63. In-Context Hallucinations
  64. Hallucinations Due to Irrelevant Information
  65. Reasoning
  66. Inducing Reasoning in LLMs
  67. Summary
  68. 9. Inference Optimization
  69. LLM Inference Challenges
  70. Inference Optimization Techniques
  71. Techniques for Reducing Compute
  72. Techniques for Accelerating Decoding
  73. Techniques for Reducing Storage Needs
  74. Summary
  75. Part III. LLM Application Paradigms
  76. 10. Interfacing LLMs with External Tools
  77. LLM Interaction Paradigms
  78. Defining Agents
  79. Agentic Workflow
  80. Components of an Agentic System
  81. Summary
  82. 11. Representation Learning and Embeddings
  83. Introduction to Embeddings
  84. Semantic Search
  85. Similarity Measures
  86. Fine-Tuning Embedding Models
  87. Instruction Embeddings
  88. Optimizing Embedding Size
  89. Chunking
  90. Vector Databases
  91. Interpreting Embeddings
  92. Summary
  93. 12. Retrieval-Augmented Generation
  94. The Need for RAG
  95. Typical RAG Scenarios
  96. Deciding When to Retrieve
  97. The RAG Pipeline
  98. RAG for Memory Management
  99. RAG for Selecting In-Context Training Examples
  100. RAG for Model Training
  101. Limitations of RAG
  102. RAG Versus Long Context
  103. RAG Versus Fine-Tuning
  104. Summary
  105. 13. Design Patterns and System Architecture
  106. Multi-LLM Architectures
  107. Programming Paradigms
  108. Summary
  109. Index
书名：设计大语言模型应用（影印版）
作者：Suhas Pai 著
国内出版社：东南大学出版社
出版时间：2025年07月
页数：343
书号：978-7-5766-2007-8
原版书书名：Designing Large Language Model Applications
原版书出版商：O'Reilly Media
Suhas Pai
 
Suhas Pai是Hudson Labs的联合创始人、首席技术官、机器学习研究负责人。Hudson Labs是一家由Y Combinator投资的AI与金融科技初创公司。他参与了多个开源LLM的开发工作，其中包括BigScience的BLOOM LLM项目，并在该项目中担任隐私工作组的联合负责人。
 
 
The animal on the cover of Designing Large Language Model Applications is the sei whale (Balaenoptera borealis), one of the largest species of baleen whales. Weighing as much as 28 tons and growing up to 64 feet in length, they are the third-largest species of baleen whale after the blue and fin whales. Even with its large size, it is a relatively fast swimmer, reaching speeds of up to 34 miles per hour.
Sei whales can be found all over the world in both subpolar and subtropical waters. They are dark blue/gray in color with a white underside and a hook-shaped dorsal fin about two-thirds down their back. Their skin is often covered in circular scars caused by cookiecutter sharks, which are known to feed on larger animals and leave these types of marks.
Instead of teeth, sei whales have between 200 and 400 baleen plates they use to eat about 2,000 pounds of food per day. Baleen plates are hair-thin, fringe-looking sheets of keratin (the same material as fingernails) that hang from the roof of the mouth and trap prey. Sei whales are filter feeders, which means they swim with their mouths open in areas with lots of prey (typically small fish, plankton, and squid) trapping food and water in their mouths. They then push out the excess water, leaving only their food.
购买选项
定价：158.00元
书号：978-7-5766-2007-8
出版社：东南大学出版社
联系出版社邮购

© 2026，奥莱利技术咨询(北京)有限公司。版权所有。
传真: +86-10-88097463
 京ICP备05003502号