快速部署大模型 — ChatGPT、嵌入式、微調和多模態 AI 的策略與最佳實踐, 2/e
[美]斯楠·奧茲德米爾（Sinan Ozdemir） 著，馮磊 周慧梅 譯
- 出版商: 清華大學
- 出版日期: 2026-06-01
- 售價: $648
- 語言: 簡體中文
- ISBN: 7302717184
- ISBN-13: 9787302717188
-
相關分類:
Large language model
- 此書翻譯自: Quick Start Guide to Large Language Models: Strategies and Best Practices for Chatgpt, Embeddings, Fine-Tuning, and Multimodal AI, 2/e (Paperback)
-
相關翻譯:
LLM 核心攻略制霸生成式 AI：ChatGPT、嵌入技術、微調與多模態 AI 最佳實踐 (繁中版)
下單後立即進貨 (約4週~6週)
商品描述
"本書是大語言模型（LLM）領域的系統性實戰指南，由兼具理論數學背景與 AI 創業經驗的 Sinan Ozdemir 撰寫，旨在幫助讀者既理解 LLM 的底層原理，又能快速將其應用於實際項目。 全書分為四大部分共12章。第一部分（第1~4章）從 LLM 基本概念出發，系統介紹 Transformer 架構、語義搜索、提示工程等基礎知識，並通過檢索增強生成（RAG）和AI智能體案例展示實際應用。第二部分（第5~8章）深入探討模型微調、高級提示工程、自定義嵌入與模型架構以及AI對齊原則，幫助讀者從“能用”邁向“用好”。第三部分（第9~12章）聚焦前沿實踐，涵蓋多模態模型、基於人類反饋的強化學習（RLHF）、開源模型高級微調、生產部署優化及 LLM 評估方法論。第四部分（附錄A~附錄D）提供常見問題解答、術語表和應用架構參考。 本書**的特色在於理論與實踐的深度結合——每個核心概念都配有直觀的類比解釋和可運行的 Python 代碼示例，讀者可跟隨實戰項目逐步掌握從原型到生產的完整技能鏈。適合有一定編程基礎的開發者、數據科學家及希望系統掌握 LLM 技術的 AI 從業者閱讀。 "
作者簡介
Sinan Ozdemir是一名數據科學家、創業者和教育工作者。Sinan的學術生涯在約翰·霍普金斯大學（The Johns Hopkins University）渡過，主修數學專業。隨後他從事教育事業，曾經在約翰·霍普金斯大學和General Assembly公司舉辦多次數據科學講座。在此之後，他創立了旨在通過人工智能技術和數據科學力量幫助企業銷售團隊的創業公司Legion Analytics。
目錄大綱
目錄
第一部分 大語言模型導論
第 1 章 大語言模型概述····························· 2
1.1 什麼是大語言模型························· 3
LLM 的定義································· 5
1.2 流行的現代 LLM··························· 6
1.2.1 BERT ································ 6
1.2.2 GPT 家族與 ChatGPT ············ 7
1.2.3 T5 ···································· 8
1.2.4 Llama ································ 8
1.3 LLM 的關鍵特征··························· 9
1.4 理解上下文的重要性····················· 12
1.5 LLM 的工作原理·························· 12
1.5.1 預訓練階段························ 13
1.5.2 遷移學習··························· 15
1.5.3 微調································· 15
1.5.4 註意力······························ 16
1.5.5 嵌入································· 18
1.5.6 標記化······························ 18
1.5.7 超越語言建模：模型對齊與基於人類反饋的強化學習 ···························· 21
1.5.8 領域特定大語言模型············ 22
1.6 LLM 的應用································ 23
1.6.1 經典 NLP 任務···················· 23
1.6.2 自由文本生成····················· 26
1.6.3 信息檢索 / 神經語義搜索······· 27
1.6.4 聊天機器人························ 28
1.7 總結·········································· 29
第 2 章 使用 LLM 進行語義搜索 ·····························30
2.1 引言·········································· 30
2.2 任務背景···································· 31
非對稱語義搜索··························· 31
2.3 解決方案概覽······························ 33
2.4 核心組件···································· 34
2.4.1 文本嵌入器························ 34
2.4.2 如何判斷文本片段的“相似性” ····························· 34
2.4.3 文檔分塊··························· 38
2.4.4 向量數據庫························ 43
2.4.5 重新排序檢索結果··············· 44
2.4.6 API ·································· 45
2.5 整合全局：讓一切運轉起來············ 46
性能評估···································· 47
2.6 閉源組件的成本··························· 50
2.7 總結·········································· 50
第 3 章 提示工程入門······························· 51
3.1 引言·········································· 51
3.2 提示工程···································· 51
3.2.1 語言模型中的對齊··············· 52
3.2.2 直接詢問··························· 53
3.2.3 當“直接詢問”不再奏效······ 55
3.2.4 少樣本學習························ 55
3.2.5 輸出格式化························ 56
3.2.6 人設提示··························· 57
3.2.7 思維鏈提示························ 58
3.2.8 示例：基礎算術·················· 59
3.3 跨模型使用提示··························· 59
3.3.1 聊天模型補全模型··············· 59
3.3.2 Cohere 的 Command 系列 ······ 61
3.3.3 開源模型的提示工程············ 62
3.4 總結·········································· 64
第 4 章 AI 生態系統：整合各個組件············65
4.1 引言·········································· 65
4.2 閉源 AI 的性能漂移 ······················ 65
4.3 AI 的推理與思考之別 ···················· 67
4.4 案例研究 1：檢索增強生成（ RAG）·· 68
4.4.1 組件協作：檢索器與生成器··· 69
4.4.2 評估 RAG 系統··················· 74
4.5 案例研究 2：自動化 AI 智能體 ········ 76
4.5.1 思考 → 行動 → 觀察 → 響應 · 76
4.5.2 評估 AI 智能體 ··················· 81
4.6 總結·········································· 82
第二部分 充分發揮 LLM 的價值
第 5 章 利用定制微調優化 LLM ·················84
5.1 引言·········································· 84
5.2 遷移學習與微調：入門指南············ 85
5.2.1 微調流程詳解····················· 86
5.2.2 以閉源預訓練模型為基礎······ 88
5.3 OpenAI 微調 API 一覽 ··················· 88
5.3.1 OpenAI 微調 API················· 88
5.3.2 案例研究：應用評論情感分類 88
5.3.3 數據準則和最佳實踐············ 89
5.4 使用 OpenAI CLI 準備自定義示例 ···· 90
5.5 設置 OpenAI CLI·························· 92
超參數選擇與優化························ 93
5.6 我們的第一個微調 LLM················· 93
5.6.1 使用定量指標評估微調模型··· 94
5.6.2 定性評估技術····················· 97
5.6.3 將微調後的 OpenAI 模型集成到應用中·················· 100
5.6.4 OpenAI 對決開源自編碼 BERT ························ 100
5.7 總結········································ 102
第 6 章 高級提示工程····························· 103
6.1 引言········································ 103
6.2 提示註入攻擊···························· 103
6.3 輸入 / 輸出驗證 ························· 105
示例：使用 NLI 構建驗證管道 ······ 106
6.4 批量提示·································· 108
6.5 提示鏈····································· 109
6.5.1 使用提示鏈防止提示堆砌·····111
6.5.2 示例：使用多模態 LLM 進行安全鏈式操作····················113
6.6 案例研究： AI 的數學有多好···········115
6.6.1 我們的數據集： MathQA·······115
6.6.2 展示你的計算過程？測試思維鏈····················117
6.6.3 用少樣本示例鼓勵 LLM······ 120
6.6.4 示例重要嗎？重訪語義搜索· 121
6.6.5 總結 MathQA 數據集的結果· 122
6.7 總結········································ 124
第 7 章 定制嵌入與模型架構···················· 125
7.1 引言········································ 125
7.2 案例研究：構建推薦系統············· 125
7.2.1 設置問題和數據················ 126
7.2.2 定義推薦問題··················· 127
7.2.3 推薦系統的宏觀視角·········· 129
7.2.4 生成自定義描述字段以比較物品················ 132
7.2.5 使用基礎嵌入器設定基線···· 133
7.2.6 準備微調數據··················· 134
7.2.7 結果總結························· 138
7.3 總結········································ 141
第 8 章 AI 對齊的第一性原理··················· 142
8.1 引言········································ 142
8.2 對齊的對象與目的······················ 142
8.2.1 指令對齊························· 142
8.2.2 行為對齊························· 143
8.2.3 風格對齊························· 145
8.2.4 價值對齊························· 146
8.3 對齊作為偏見緩解器··················· 147
8.4 對齊的三大支柱························· 151
8.4.1 數據······························· 151
8.4.2 訓練 / 調優模型 ················ 154
8.4.3 評估······························· 156
8.4.4 對齊的三大支柱小結·········· 165
8.5 憲法式 AI：邁向自我對齊的一步 ··· 166
8.6 總結········································ 168
第三部分 LLM 高級應用
第 9 章 超越基礎模型····························· 170
9.1 引言········································ 170
9.2 案例研究：視覺問答··················· 170
9.2.1 模型介紹：Vision Transformer、 GPT-2 和 DistilBERT····························· 171
9.2.2 隱藏狀態投影與融合·········· 175
9.2.3 交叉註意力：它是什麼？為什麼如此關鍵·························· 175
9.2.4 我們的自定義多模態模型···· 178
9.2.5 我們的數據： Visual QA······· 180
9.2.6 VQA 訓練循環·················· 182
9.2.7 結果總結························· 183
9.3 案例研究：基於反饋的強化學習···· 185
9.3.1 我們的模型： FLAN-T5 ······· 186
9.3.2 我們的獎勵模型：情感與語法正確性··················· 187
9.3.3 Transformer 強化學習 ········· 188
9.3.4 RLF 訓練循環 ·················· 189
9.3.5 結果總結························· 192
9.4 總結········································ 194
第 10 章 高級開源 LLM 微調··················· 195
10.1 引言 ······································ 195
10.2 示例：使用 BERT 進行動漫流派多標簽分類 ····················· 195
10.2.1 使用 Jaccard 分數衡量動漫標題多標簽流派預測的性能························ 196
10.2.2 一個簡單的微調循環 ········ 197
10.2.3 微調開源 LLM 的通用技巧 199
10.2.4 結果總結 ······················· 205
10.3 示例：使用 GPT2 生成 LaTeX······ 207
10.3.1 開源模型的提示工程 ········ 208
10.3.2 結果總結 ······················· 210
10.4 打造自己的睿智且引人入勝的對話助手—SAWYER ···························211
10.4.1 步驟 1：監督指令微調······ 213
10.4.2 步驟 2：獎勵模型訓練······ 218
10.4.3 步驟 3：基於（模擬的）人類反饋的強化學習··································· 222
10.4.4 結果總結 ······················· 225
10.4.5 用新鮮知識更新我們的 LLM ·························· 228
10.5 總結 ······································ 230
第 11 章 將 LLM 投入生產······················ 233
11.1 引言 ······································ 233
11.2 閉源 LLM 的生產部署················ 233
11.2.1 成本預估 ······················· 233
11.2.2 API 密鑰管理·················· 234
11.3 開源 LLM 的生產部署················ 234
11.3.1 為推理準備模型 ·············· 234
11.3.2 互操作性 ······················· 234
11.3.3 量化 ····························· 235
11.3.4 知識蒸餾 ······················· 240
11.3.5 LLM 的成本預估 ············· 248
11.3.6 發布到 Hugging Face········· 249
11.4 總結 ······································ 252
第 12 章 評估 LLM ······························· 253
12.1 引言 ······································ 253
12.2 評估生成任務 ·························· 254
12.2.1 生成式多項選擇 ·············· 254
12.2.2 自由文本回答 ················· 257
12.2.3 基準測試 ······················· 259
12.3 評估理解任務 ·························· 267
12.3.1 嵌入 ····························· 267
12.3.2 校準分類 ······················· 270
12.3.3 探測 LLM 的世界模型 ······ 273
12.4 總結 ······································ 277
12.5 繼續前行 ································ 278
第四部分 附錄
附錄 A LLM 常見問題解答 ····················· 280
附錄 B LLM 術語表 ······························ 283
附錄 C LLM 應用原型 ··························· 288
附錄 D 代碼倉庫使用指南······················· 291
