├── subtitle_split.py ├── chinese_subtitles ├── prompt_eng_09_conclusion_chinese.srt ├── prompt_eng_04_summarizing_chinese.srt ├── prompt_eng_01_intro_chinese.srt ├── prompt_eng_07_expanding_chinese.srt ├── prompt_eng_06_transforming_chinese.srt ├── prompt_eng_05_inferring_chinese.srt └── prompt_eng_08_chatbot_chinese.srt ├── README.md ├── english_subtitles ├── prompt_eng_09_conclusion_english.srt ├── prompt_eng_04_summarizing_english.srt ├── prompt_eng_01_intro_english.srt └── prompt_eng_07_expanding_english.srt ├── subtitles ├── raw │ ├── prompt_eng_09_conclusion.srt │ ├── prompt_eng_04_summarizing.srt │ └── prompt_eng_01_intro.srt └── styled │ └── prompt_eng_09_conclusion.ass └── course_notebooks ├── l7-expanding.ipynb ├── l8-chatbot.ipynb ├── l3-iterative-prompt-development.ipynb ├── l5-inferring.ipynb ├── l6-transforming.ipynb └── l4-summarizing.ipynb /subtitle_split.py: -------------------------------------------------------------------------------- 1 | import os 2 | import re 3 | 4 | subs_path = 'subtitles/raw' 5 | cn_subs_path = 'chinese_subtitles' 6 | en_subs_path = 'english_subtitles' 7 | 8 | def extract_timeline_and_content(input_file): 9 | # Read the input file and create a dictionary of subtitles with timeline keys 10 | with open(input_file, "r", encoding="utf-8") as f: 11 | lines = f.readlines() 12 | 13 | subtitles_dict = {} 14 | for i, line in enumerate(lines): 15 | line = line.strip() 16 | if re.match(r'\d{2}:\d{2}:\d{2},\d{3} --> \d{2}:\d{2}:\d{2},\d{3}', line): 17 | timeline = line 18 | chinese = lines[i + 1].strip() 19 | english = lines[i + 2].strip() 20 | subtitles_dict[timeline] = {"chinese": chinese, "english": english} 21 | 22 | return subtitles_dict 23 | 24 | def write_chinese_subtitles(subtitles_dict, output_file): 25 | # Write Chinese subtitles to output_file 26 | with open(output_file, "w", encoding="utf-8") as f: 27 | for idx, (timeline, content) in enumerate(subtitles_dict.items()): 28 | f.write(f"{idx + 1}\n{timeline}\n{content['chinese']}\n\n") 29 | 30 | def write_english_subtitles(subtitles_dict, output_file): 31 | # Write English subtitles to output_file 32 | with open(output_file, "w", encoding="utf-8") as f: 33 | for idx, (timeline, content) in enumerate(subtitles_dict.items()): 34 | f.write(f"{idx + 1}\n{timeline}\n{content['english']}\n\n") 35 | 36 | def process_filename(filename): 37 | # Generate Chinese and English output filenames 38 | basename, ext = os.path.splitext(filename) 39 | return basename + "_chinese" + ext, basename + "_english" + ext 40 | 41 | def process_files(): 42 | # Create output directories if they don't exist 43 | os.makedirs(cn_subs_path, exist_ok=True) 44 | os.makedirs(en_subs_path, exist_ok=True) 45 | 46 | # Iterate through all SRT files in the subtitles directory 47 | for filename in os.listdir(subs_path): 48 | if filename.endswith(".srt") and not filename.endswith("_chinese.srt") and not filename.endswith("_english.srt"): 49 | print(f"Processing {filename}") 50 | input_file = os.path.join(subs_path, filename) 51 | subtitles_dict = extract_timeline_and_content(input_file) 52 | chinese_output, english_output = process_filename(filename) 53 | write_chinese_subtitles(subtitles_dict, os.path.join(cn_subs_path, chinese_output)) 54 | write_english_subtitles(subtitles_dict, os.path.join(en_subs_path, english_output)) 55 | print(f"Generated {chinese_output} and {english_output}\n") 56 | 57 | # Traverse all subtitle files under current directory and generate Chinese and English subtitles. 58 | process_files() -------------------------------------------------------------------------------- /chinese_subtitles/prompt_eng_09_conclusion_chinese.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:08,840 3 | 祝贺你在这个短期课程中走到了最后 4 | 5 | 2 6 | 00:00:08,840 --> 00:00:11,280 7 | 综上所述，在这个短课程中 8 | 9 | 3 10 | 00:00:11,280 --> 00:00:14,380 11 | 你已经了解了提示的两个关键原则 12 | 13 | 4 14 | 00:00:14,380 --> 00:00:17,480 15 | 正确且清晰的描述具体指示，以及知道什么时候使用 16 | 17 | 5 18 | 00:00:17,480 --> 00:00:19,160 19 | 给予模型思考的时间 20 | 21 | 6 22 | 00:00:19,160 --> 00:00:22,080 23 | 你还了解了迭代提示的发展以及如何 24 | 25 | 7 26 | 00:00:22,080 --> 00:00:26,200 27 | 有一个过程来获得适合你的应用的提示是关键 28 | 29 | 8 30 | 00:00:26,200 --> 00:00:28,800 31 | 我们还学习了一些大语言模型的功能 32 | 33 | 9 34 | 00:00:28,800 --> 00:00:31,640 35 | 这些功能对许多应用是有用的 36 | 37 | 10 38 | 00:00:31,640 --> 00:00:36,320 39 | 特别是总结、推理、转换和扩展 40 | 41 | 11 42 | 00:00:36,320 --> 00:00:39,880 43 | 你还看到了如何建立一个定制的聊天机器人 44 | 45 | 12 46 | 00:00:39,880 --> 00:00:43,320 47 | 在短短的课程中，你学到了很多东西 48 | 49 | 13 50 | 00:00:43,320 --> 00:00:46,240 51 | 我希望你喜欢阅读这些材料 52 | 53 | 14 54 | 00:00:46,240 --> 00:00:47,920 55 | 我们希望你能想出一些想法 56 | 57 | 15 58 | 00:00:47,920 --> 00:00:50,120 59 | 我们希望你能想出一些可以自己构建的应用 60 | 61 | 16 62 | 00:00:50,120 --> 00:00:53,120 63 | 并请你去尝试一下，让我们知道你想出了什么 64 | 65 | 17 66 | 00:00:53,120 --> 00:00:54,720 67 | 任何应用都不小 68 | 69 | 18 70 | 00:00:54,720 --> 00:00:57,000 71 | 从一些很小的项目开始也是可以的 72 | 73 | 19 74 | 00:00:57,000 --> 00:00:59,800 75 | 一个非常小的项目，也许实用 76 | 77 | 20 78 | 00:00:59,800 --> 00:01:01,800 79 | 也可能毫无用处 80 | 81 | 21 82 | 00:01:01,800 --> 00:01:03,320 83 | 这只是一些有趣的东西 84 | 85 | 22 86 | 00:01:03,320 --> 00:01:06,200 87 | 是的。我发现玩这些模型非常有趣 88 | 89 | 23 90 | 00:01:06,200 --> 00:01:07,440 91 | 所以放心去玩吧 92 | 93 | 24 94 | 00:01:07,440 --> 00:01:08,640 95 | 是的，我同意。从我的经验来看 96 | 97 | 25 98 | 00:01:08,640 --> 00:01:11,080 99 | 这是一个很好的周末活动 100 | 101 | 26 102 | 00:01:11,080 --> 00:01:14,960 103 | 而且，你可以通过第一个项目获得的经验教训 104 | 105 | 27 106 | 00:01:14,960 --> 00:01:18,520 107 | 来构建第二个更好的项目 108 | 109 | 28 110 | 00:01:18,520 --> 00:01:21,360 111 | 甚至可能是更好的第三个项目，等等 112 | 113 | 29 114 | 00:01:21,360 --> 00:01:26,960 115 | 这就是我自己使用这些模型逐渐成长的方式 116 | 117 | 30 118 | 00:01:26,960 --> 00:01:31,360 119 | 或者如果你已经有了一个更大的项目想法，就去做吧 120 | 121 | 31 122 | 00:01:31,360 --> 00:01:33,960 123 | 作为一个提醒 124 | 125 | 32 126 | 00:01:33,960 --> 00:01:37,200 127 | 这些大语言模型是一种非常强大的技术 128 | 129 | 33 130 | 00:01:37,200 --> 00:01:41,120 131 | 所以不言而喻，我们要求你负责任地使用它们 132 | 133 | 34 134 | 00:01:41,120 --> 00:01:43,960 135 | 请只建立会产生积极影响的东西 136 | 137 | 35 138 | 00:01:43,960 --> 00:01:46,240 139 | 是的，完全同意。我认为在这个时代 140 | 141 | 36 142 | 00:01:46,240 --> 00:01:49,480 143 | 建立AI系统的人可以对其他人产生巨大的影响 144 | 145 | 37 146 | 00:01:49,480 --> 00:01:54,960 147 | 因此，我们所有人都要负责任地使用这些工具，这一点比以往任何时候都重要 148 | 149 | 38 150 | 00:01:54,960 --> 00:01:59,720 151 | 我认为建立基于大语言模型的应用程序 152 | 153 | 39 154 | 00:01:59,720 --> 00:02:02,880 155 | 是一个非常令人兴奋和不断增长的领域 156 | 157 | 40 158 | 00:02:02,880 --> 00:02:05,160 159 | 现在你已经完成了这个课程 160 | 161 | 41 162 | 00:02:05,160 --> 00:02:08,240 163 | 我想你现在已经有了丰富的知识 164 | 165 | 42 166 | 00:02:08,240 --> 00:02:11,320 167 | 可以让你构建少数人知道如何构建的东西 168 | 169 | 43 170 | 00:02:11,320 --> 00:02:17,280 171 | 与此同时，我希望你也能帮助我们传播信息，鼓励其他人也参加这个课程 172 | 173 | 44 174 | 00:02:17,280 --> 00:02:20,520 175 | 最后，我希望你在做这个课程时很开心 176 | 177 | 45 178 | 00:02:20,520 --> 00:02:23,280 179 | 同时也感谢你能完成这个课程 180 | 181 | 46 182 | 00:02:23,280 --> 00:02:27,120 183 | Isa和我都期待着听到你所建立的惊人的东西 184 | 185 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | ![banner](https://user-images.githubusercontent.com/42373389/235295056-fc1d7ba7-112d-45e8-a56f-b8cf68d9fa31.png) 2 | 3 | # 《ChatGPT Prompt Engineering for Developers》中英双语字幕 4 | 5 | 该项目主要用于存放《面向开发者的 ChatGPT 提示词工程》非官方版中英双语字幕。 6 | 7 | - **中文视频地址：[面向开发者的 ChatGPT 提示词工程](https://www.bilibili.com/video/BV1s24y1F7eq/)** 8 | - **英文原视频地址：[ChatGPT Prompt Engineering for Developers](https://learn.deeplearning.ai/)** 9 | - **配套代码与笔记：[《ChatGPT Prompt Engineering for Developers》课程中文版 Datawhale](https://github.com/datawhalechina/prompt-engineering-for-developers)** 10 | 11 | 如果你在观看视频的过程中，发现翻译出现错漏、错别字、病句等情况，欢迎向我们提交 Pull Request 以改进字幕翻译质量。 12 | 13 | 本项目文件夹说明： 14 | 15 | - `subtitles`：核心字幕，里面放置了「中英双语字幕」； 16 | - `english_subtitles`：从双语字幕中分离出来的「英文字幕」； 17 | - `chinese_subtitles`：从双语字幕中分离出来的「中文字幕」; 18 | - `course_notebooks`: 从课程中复制至本地的 Notebook。 19 | 20 | 字幕效果展示： 21 | 22 | ![image](https://user-images.githubusercontent.com/42373389/235338205-31271b4c-b379-43d1-9b3e-bd5114f9dc09.png) 23 | 24 | ## 课程介绍 25 | 26 | ChatGPT 上线至今，已经快 5 个月了，但是不少人还没真正掌握它的使用技巧。 27 | 28 | 其实，ChatGPT 的难点，在于 Prompt（提示词）的编写，OpenAI 创始人在今年 2 月时，在 Twitter 上说：「能够出色编写 Prompt 跟聊天机器人对话，是一项能令人惊艳的高杠杆技能」。 29 | 30 | 因为从 ChatGPT 发布之后，如何写好 Prompt 已经成为了一个分水岭。熟练掌握 Prompt 编写的人，能够很快让 ChatGPT 理解需求，并很好的执行任务。 31 | 32 | 目前你在网上看到的所有 AI 助理、智能翻译、角色扮演，本质上还是通过编写 Prompt 来实现。 33 | 34 | 只要你的 Prompt 写的足够好，ChatGPT 可以帮你快速完成很多工作，包括写爬虫脚本、金融数据分析、文案润色与翻译等等，并且这些工作还做的比一般人出色。 35 | 36 | 为了帮助大家能更好的掌握 Prompt 工程，DeepLearning.ai 创始人吴恩达与 OpenAI 开发者 Iza Fulford 联手推出了一门面向开发者的技术教程：《**ChatGPT 提示工程**》。 37 | 38 |

39 | 40 |

41 | 42 | 吴恩达老师相信大家都有所耳闻，作为人工智能界的重量级大佬，我们经常能在 AI 技术界看到他活跃的身影。 43 | 44 | 另一位讲师 Iza Fulford，大家可能不太熟悉，这里重点介绍下。 45 | 46 | 她是斯坦福本硕高材生，ChatGPT 之前在 GitHub 开源的那个文档搜索插件：**Retrieval**，就是出自她之手。 47 | 48 | 另外，她还是 OpenAI Cookbook（官方手册）的编撰者，如果你最近有深入了解过 GPT 相关的技术，那这本手册于你而言应该不会陌生。 49 | 50 | 该手册里面提供了大量 GPT 相关的使用案例，能帮助你快速上手并掌握 GPT 模型的开发与应用。 51 | 52 | 可以说，这两位大佬联手，推出的教程绝对不会差。更令人振奋的是，这个教程完全对外开放，所有人均可免费学习！ 53 | 54 | 那么，这个教程里面主要讲了什么内容呢？ 55 | 56 | 该教程总共分为 9 个章节，总一个多小时，里面主要涵盖：**提示词最佳实践、评论情感分类、文本总结、邮件撰写、文本翻译、快速搭建一个聊天机器人**等等。 57 | 58 | ![image](https://user-images.githubusercontent.com/42373389/235334065-97349f08-ac8b-41c5-ad1a-b2ed0bec7f3f.png) 59 | 60 | 所有当下 ChatGPT 的流行案例，你都能在这个教程里面找到，十分全面！ 61 | 62 | 除了能在这个教程里面学到如何使用 Prompt，你还能学到 GPT 接口调用开发知识。有需要的话，你甚至能在这个教程之上去延伸扩展，搭建出一款令人惊艳的应用。 63 | 64 | 目前该教程已经在 DeepLearning.ai 正式上线，官网上线提供了可交互式的 Notebook，让你可以一边学习，一边跟着编写代码实践。 65 | 66 | 不过当下这个教程只有英文版，为了让看不懂英文的同学也能第一时间学习并掌握这项技术。 67 | 68 | 我花了一天时间，完整翻译了所有英文字幕，并且[将所有视频与字幕同步上传到了 B 站](https://www.bilibili.com/video/BV1s24y1F7eq/)。 69 | 70 | 大家有时间的话，可以多学一下这个教程，相信整个学习下来，绝对能让你受益匪浅！ 71 | 72 | ## 贡献者 73 | 74 | 查看 [贡献列表](https://github.com/GitHubDaily/ChatGPT-Prompt-Engineering-for-Developers-in-Chinese/graphs/contributors) 获取更多信息，感谢所有为项目作为贡献的人！ 75 | 76 | 77 | 78 | 79 | 80 | ## 开源协议 81 | 82 | 项目基于 [CC BY-NC-SA 4.0 协议](https://creativecommons.org/licenses/by-nc-sa/4.0/deed.zh) 发布。 83 | 84 | [![CC BY-NC-SA 4.0][cc-by-nc-sa-image]][cc-by-nc-sa] 85 | 86 | [cc-by-nc-sa]: http://creativecommons.org/licenses/by-nc-sa/4.0/ 87 | [cc-by-nc-sa-image]: https://licensebuttons.net/l/by-nc-sa/4.0/88x31.png 88 | [cc-by-nc-sa-shield]: https://img.shields.io/badge/License-CC%20BY--NC--SA%204.0-lightgrey.svg 89 | -------------------------------------------------------------------------------- /english_subtitles/prompt_eng_09_conclusion_english.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:08,840 3 | Congratulations on making it to the end of this short course 4 | 5 | 2 6 | 00:00:08,840 --> 00:00:11,280 7 | In summary, in this short course 8 | 9 | 3 10 | 00:00:11,280 --> 00:00:14,380 11 | you've learned about two key principles for prompting 12 | 13 | 4 14 | 00:00:14,380 --> 00:00:17,480 15 | right clearance for specific instructions and when it's appropriate 16 | 17 | 5 18 | 00:00:17,480 --> 00:00:19,160 19 | give the model time to think 20 | 21 | 6 22 | 00:00:19,160 --> 00:00:22,080 23 | You also learned about iterative prompt development and how 24 | 25 | 7 26 | 00:00:22,080 --> 00:00:26,200 27 | having a process to get to the prompt is right for your application is key 28 | 29 | 8 30 | 00:00:26,200 --> 00:00:28,800 31 | And we went through a few capabilities of 32 | 33 | 9 34 | 00:00:28,800 --> 00:00:31,640 35 | large language models that are useful for many applications 36 | 37 | 10 38 | 00:00:31,640 --> 00:00:36,320 39 | specifically summarizing, inferring, transforming, and expanding 40 | 41 | 11 42 | 00:00:36,320 --> 00:00:39,880 43 | And you also saw how to build a custom chatbot 44 | 45 | 12 46 | 00:00:39,880 --> 00:00:43,320 47 | That was a lot that you learned in just one short course 48 | 49 | 13 50 | 00:00:43,320 --> 00:00:46,240 51 | and I hope you enjoyed going through these materials 52 | 53 | 14 54 | 00:00:46,240 --> 00:00:47,920 55 | We hope you'll come up with some ideas for 56 | 57 | 15 58 | 00:00:47,920 --> 00:00:50,120 59 | applications that you can build yourself now 60 | 61 | 16 62 | 00:00:50,120 --> 00:00:53,120 63 | And please go try this out and let us know what you come up with 64 | 65 | 17 66 | 00:00:53,120 --> 00:00:54,720 67 | No application is too small 68 | 69 | 18 70 | 00:00:54,720 --> 00:00:57,000 71 | It's fine to start with something that's kind of 72 | 73 | 19 74 | 00:00:57,000 --> 00:00:59,800 75 | a very small project with maybe a little bit of utility 76 | 77 | 20 78 | 00:00:59,800 --> 00:01:01,800 79 | or maybe it's not even useful at all 80 | 81 | 21 82 | 00:01:01,800 --> 00:01:03,320 83 | It's just something fun 84 | 85 | 22 86 | 00:01:03,320 --> 00:01:06,200 87 | Yeah. And I find playing with these models actually really fun 88 | 89 | 23 90 | 00:01:06,200 --> 00:01:07,440 91 | So go play with it 92 | 93 | 24 94 | 00:01:07,440 --> 00:01:08,640 95 | Yeah, I agree. It's a good 96 | 97 | 25 98 | 00:01:08,640 --> 00:01:11,080 99 | a good weekend activity, speaking from experience 100 | 101 | 26 102 | 00:01:11,080 --> 00:01:14,960 103 | And just, you know 104 | 105 | 27 106 | 00:01:14,960 --> 00:01:18,520 107 | please use the learnings from your first project to build a better second project 108 | 109 | 28 110 | 00:01:18,520 --> 00:01:21,360 111 | and you know, maybe even a better third project, so on 112 | 113 | 29 114 | 00:01:21,360 --> 00:01:26,960 115 | That's kind of how I have kind of grown over time using these models myself as well 116 | 117 | 30 118 | 00:01:26,960 --> 00:01:31,360 119 | Or if you have an idea for a bigger project already, just, just go for it 120 | 121 | 31 122 | 00:01:31,360 --> 00:01:33,960 123 | And, you know, as a reminder 124 | 125 | 32 126 | 00:01:33,960 --> 00:01:37,200 127 | these kind of large language models are a very powerful technology 128 | 129 | 33 130 | 00:01:37,200 --> 00:01:41,120 131 | So it kind of goes without saying that we ask you to use them responsibly 132 | 133 | 34 134 | 00:01:41,120 --> 00:01:43,960 135 | and please only build things that will have a positive impact 136 | 137 | 35 138 | 00:01:43,960 --> 00:01:46,240 139 | Yeah, fully agree. I think in this age 140 | 141 | 36 142 | 00:01:46,240 --> 00:01:49,480 143 | people that build AI systems can have a huge impact on others 144 | 145 | 37 146 | 00:01:49,480 --> 00:01:54,960 147 | So it's more important than ever that all of us only use these tools responsibly 148 | 149 | 38 150 | 00:01:54,960 --> 00:01:59,720 151 | And I think building large language model-based applications 152 | 153 | 39 154 | 00:01:59,720 --> 00:02:02,880 155 | is just a very exciting and growing field right now 156 | 157 | 40 158 | 00:02:02,880 --> 00:02:05,160 159 | And now that you've finished this course 160 | 161 | 41 162 | 00:02:05,160 --> 00:02:08,240 163 | I think you now have a wealth of knowledge to let 164 | 165 | 42 166 | 00:02:08,240 --> 00:02:11,320 167 | you build things that few people today know how to 168 | 169 | 43 170 | 00:02:11,320 --> 00:02:17,280 171 | So I hope you also help us to spread the word and encourage others to take this course too 172 | 173 | 44 174 | 00:02:17,280 --> 00:02:20,520 175 | In closing, I hope you had fun doing this course 176 | 177 | 45 178 | 00:02:20,520 --> 00:02:23,280 179 | and I want to thank you for finishing this course 180 | 181 | 46 182 | 00:02:23,280 --> 00:02:27,120 183 | and both Isa and I look forward to hearing about the amazing things that you've built 184 | 185 | -------------------------------------------------------------------------------- /subtitles/raw/prompt_eng_09_conclusion.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:08,840 3 | 祝贺你在这个短期课程中走到了最后 4 | Congratulations on making it to the end of this short course 5 | 6 | 2 7 | 00:00:08,840 --> 00:00:11,280 8 | 综上所述，在这个短课程中 9 | In summary, in this short course 10 | 11 | 3 12 | 00:00:11,280 --> 00:00:14,380 13 | 你已经了解了提示的两个关键原则 14 | you've learned about two key principles for prompting 15 | 16 | 4 17 | 00:00:14,380 --> 00:00:17,480 18 | 正确且清晰的描述具体指示，以及知道什么时候使用 19 | right clearance for specific instructions and when it's appropriate 20 | 21 | 5 22 | 00:00:17,480 --> 00:00:19,160 23 | 给予模型思考的时间 24 | give the model time to think 25 | 26 | 6 27 | 00:00:19,160 --> 00:00:22,080 28 | 你还了解了迭代提示的发展以及如何 29 | You also learned about iterative prompt development and how 30 | 31 | 7 32 | 00:00:22,080 --> 00:00:26,200 33 | 有一个过程来获得适合你的应用的提示是关键 34 | having a process to get to the prompt is right for your application is key 35 | 36 | 8 37 | 00:00:26,200 --> 00:00:28,800 38 | 我们还学习了一些大语言模型的功能 39 | And we went through a few capabilities of 40 | 41 | 9 42 | 00:00:28,800 --> 00:00:31,640 43 | 这些功能对许多应用是有用的 44 | large language models that are useful for many applications 45 | 46 | 10 47 | 00:00:31,640 --> 00:00:36,320 48 | 特别是总结、推理、转换和扩展 49 | specifically summarizing, inferring, transforming, and expanding 50 | 51 | 11 52 | 00:00:36,320 --> 00:00:39,880 53 | 你还看到了如何建立一个定制的聊天机器人 54 | And you also saw how to build a custom chatbot 55 | 56 | 12 57 | 00:00:39,880 --> 00:00:43,320 58 | 在短短的课程中，你学到了很多东西 59 | That was a lot that you learned in just one short course 60 | 61 | 13 62 | 00:00:43,320 --> 00:00:46,240 63 | 我希望你喜欢阅读这些材料 64 | and I hope you enjoyed going through these materials 65 | 66 | 14 67 | 00:00:46,240 --> 00:00:47,920 68 | 我们希望你能想出一些想法 69 | We hope you'll come up with some ideas for 70 | 71 | 15 72 | 00:00:47,920 --> 00:00:50,120 73 | 我们希望你能想出一些可以自己构建的应用 74 | applications that you can build yourself now 75 | 76 | 16 77 | 00:00:50,120 --> 00:00:53,120 78 | 并请你去尝试一下，让我们知道你想出了什么 79 | And please go try this out and let us know what you come up with 80 | 81 | 17 82 | 00:00:53,120 --> 00:00:54,720 83 | 任何应用都不小 84 | No application is too small 85 | 86 | 18 87 | 00:00:54,720 --> 00:00:57,000 88 | 从一些很小的项目开始也是可以的 89 | It's fine to start with something that's kind of 90 | 91 | 19 92 | 00:00:57,000 --> 00:00:59,800 93 | 一个非常小的项目，也许实用 94 | a very small project with maybe a little bit of utility 95 | 96 | 20 97 | 00:00:59,800 --> 00:01:01,800 98 | 也可能毫无用处 99 | or maybe it's not even useful at all 100 | 101 | 21 102 | 00:01:01,800 --> 00:01:03,320 103 | 这只是一些有趣的东西 104 | It's just something fun 105 | 106 | 22 107 | 00:01:03,320 --> 00:01:06,200 108 | 是的。我发现玩这些模型非常有趣 109 | Yeah. And I find playing with these models actually really fun 110 | 111 | 23 112 | 00:01:06,200 --> 00:01:07,440 113 | 所以放心去玩吧 114 | So go play with it 115 | 116 | 24 117 | 00:01:07,440 --> 00:01:08,640 118 | 是的，我同意。从我的经验来看 119 | Yeah, I agree. It's a good 120 | 121 | 25 122 | 00:01:08,640 --> 00:01:11,080 123 | 这是一个很好的周末活动 124 | a good weekend activity, speaking from experience 125 | 126 | 26 127 | 00:01:11,080 --> 00:01:14,960 128 | 而且，你可以通过第一个项目获得的经验教训 129 | And just, you know 130 | 131 | 27 132 | 00:01:14,960 --> 00:01:18,520 133 | 来构建第二个更好的项目 134 | please use the learnings from your first project to build a better second project 135 | 136 | 28 137 | 00:01:18,520 --> 00:01:21,360 138 | 甚至可能是更好的第三个项目，等等 139 | and you know, maybe even a better third project, so on 140 | 141 | 29 142 | 00:01:21,360 --> 00:01:26,960 143 | 这就是我自己使用这些模型逐渐成长的方式 144 | That's kind of how I have kind of grown over time using these models myself as well 145 | 146 | 30 147 | 00:01:26,960 --> 00:01:31,360 148 | 或者如果你已经有了一个更大的项目想法，就去做吧 149 | Or if you have an idea for a bigger project already, just, just go for it 150 | 151 | 31 152 | 00:01:31,360 --> 00:01:33,960 153 | 作为一个提醒 154 | And, you know, as a reminder 155 | 156 | 32 157 | 00:01:33,960 --> 00:01:37,200 158 | 这些大语言模型是一种非常强大的技术 159 | these kind of large language models are a very powerful technology 160 | 161 | 33 162 | 00:01:37,200 --> 00:01:41,120 163 | 所以不言而喻，我们要求你负责任地使用它们 164 | So it kind of goes without saying that we ask you to use them responsibly 165 | 166 | 34 167 | 00:01:41,120 --> 00:01:43,960 168 | 请只建立会产生积极影响的东西 169 | and please only build things that will have a positive impact 170 | 171 | 35 172 | 00:01:43,960 --> 00:01:46,240 173 | 是的，完全同意。我认为在这个时代 174 | Yeah, fully agree. I think in this age 175 | 176 | 36 177 | 00:01:46,240 --> 00:01:49,480 178 | 建立AI系统的人可以对其他人产生巨大的影响 179 | people that build AI systems can have a huge impact on others 180 | 181 | 37 182 | 00:01:49,480 --> 00:01:54,960 183 | 因此，我们所有人都要负责任地使用这些工具，这一点比以往任何时候都重要 184 | So it's more important than ever that all of us only use these tools responsibly 185 | 186 | 38 187 | 00:01:54,960 --> 00:01:59,720 188 | 我认为建立基于大语言模型的应用程序 189 | And I think building large language model-based applications 190 | 191 | 39 192 | 00:01:59,720 --> 00:02:02,880 193 | 是一个非常令人兴奋和不断增长的领域 194 | is just a very exciting and growing field right now 195 | 196 | 40 197 | 00:02:02,880 --> 00:02:05,160 198 | 现在你已经完成了这个课程 199 | And now that you've finished this course 200 | 201 | 41 202 | 00:02:05,160 --> 00:02:08,240 203 | 我想你现在已经有了丰富的知识 204 | I think you now have a wealth of knowledge to let 205 | 206 | 42 207 | 00:02:08,240 --> 00:02:11,320 208 | 可以让你构建少数人知道如何构建的东西 209 | you build things that few people today know how to 210 | 211 | 43 212 | 00:02:11,320 --> 00:02:17,280 213 | 与此同时，我希望你也能帮助我们传播信息，鼓励其他人也参加这个课程 214 | So I hope you also help us to spread the word and encourage others to take this course too 215 | 216 | 44 217 | 00:02:17,280 --> 00:02:20,520 218 | 最后，我希望你在做这个课程时很开心 219 | In closing, I hope you had fun doing this course 220 | 221 | 45 222 | 00:02:20,520 --> 00:02:23,280 223 | 同时也感谢你能完成这个课程 224 | and I want to thank you for finishing this course 225 | 226 | 46 227 | 00:02:23,280 --> 00:02:27,120 228 | Isa和我都期待着听到你所建立的惊人的东西 229 | and both Isa and I look forward to hearing about the amazing things that you've built 230 | 231 | -------------------------------------------------------------------------------- /course_notebooks/l7-expanding.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "attachments": {}, 5 | "cell_type": "markdown", 6 | "metadata": {}, 7 | "source": [ 8 | "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/GitHubDaily/ChatGPT-Prompt-Engineering-for-Developers-in-Chinese/blob/master/assets/7-Expanding/l7-expanding.ipynb)\n", 9 | "# Expanding\n", 10 | "In this lesson, you will generate customer service emails that are tailored to each customer's review.\n", 11 | "\n", 12 | "## Setup" 13 | ] 14 | }, 15 | { 16 | "cell_type": "code", 17 | "execution_count": null, 18 | "metadata": {}, 19 | "outputs": [], 20 | "source": [ 21 | "# Basic congfig \n", 22 | "# Install basic package and set key\n", 23 | "!pip install openai\n", 24 | "!export OPENAI_API_KEY='sk-...'" 25 | ] 26 | }, 27 | { 28 | "cell_type": "code", 29 | "execution_count": null, 30 | "metadata": {}, 31 | "outputs": [], 32 | "source": [ 33 | "import openai\n", 34 | "import os\n", 35 | "\n", 36 | "from dotenv import load_dotenv, find_dotenv\n", 37 | "_ = load_dotenv(find_dotenv()) # read local .env file\n", 38 | "\n", 39 | "openai.api_key = os.getenv('OPENAI_API_KEY')" 40 | ] 41 | }, 42 | { 43 | "cell_type": "code", 44 | "execution_count": null, 45 | "metadata": {}, 46 | "outputs": [], 47 | "source": [ 48 | "def get_completion(prompt, model=\"gpt-3.5-turbo\",temperature=0): # Andrew mentioned that the prompt/ completion paradigm is preferable for this class\n", 49 | " messages = [{\"role\": \"user\", \"content\": prompt}]\n", 50 | " response = openai.ChatCompletion.create(\n", 51 | " model=model,\n", 52 | " messages=messages,\n", 53 | " temperature=temperature, # this is the degree of randomness of the model's output\n", 54 | " )\n", 55 | " return response.choices[0].message[\"content\"]" 56 | ] 57 | }, 58 | { 59 | "attachments": {}, 60 | "cell_type": "markdown", 61 | "metadata": {}, 62 | "source": [ 63 | "## Customize the automated reply to a customer email" 64 | ] 65 | }, 66 | { 67 | "cell_type": "code", 68 | "execution_count": null, 69 | "metadata": {}, 70 | "outputs": [], 71 | "source": [ 72 | "# given the sentiment from the lesson on \"inferring\",\n", 73 | "# and the original customer message, customize the email\n", 74 | "sentiment = \"negative\"\n", 75 | "\n", 76 | "# review for a blender\n", 77 | "review = f\"\"\"\n", 78 | "So, they still had the 17 piece system on seasonal \\\n", 79 | "sale for around $49 in the month of November, about \\\n", 80 | "half off, but for some reason (call it price gouging) \\\n", 81 | "around the second week of December the prices all went \\\n", 82 | "up to about anywhere from between $70-$89 for the same \\\n", 83 | "system. And the 11 piece system went up around $10 or \\\n", 84 | "so in price also from the earlier sale price of $29. \\\n", 85 | "So it looks okay, but if you look at the base, the part \\\n", 86 | "where the blade locks into place doesn’t look as good \\\n", 87 | "as in previous editions from a few years ago, but I \\\n", 88 | "plan to be very gentle with it (example, I crush \\\n", 89 | "very hard items like beans, ice, rice, etc. in the \\ \n", 90 | "blender first then pulverize them in the serving size \\\n", 91 | "I want in the blender then switch to the whipping \\\n", 92 | "blade for a finer flour, and use the cross cutting blade \\\n", 93 | "first when making smoothies, then use the flat blade \\\n", 94 | "if I need them finer/less pulpy). Special tip when making \\\n", 95 | "smoothies, finely cut and freeze the fruits and \\\n", 96 | "vegetables (if using spinach-lightly stew soften the \\ \n", 97 | "spinach then freeze until ready for use-and if making \\\n", 98 | "sorbet, use a small to medium sized food processor) \\ \n", 99 | "that you plan to use that way you can avoid adding so \\\n", 100 | "much ice if at all-when making your smoothie. \\\n", 101 | "After about a year, the motor was making a funny noise. \\\n", 102 | "I called customer service but the warranty expired \\\n", 103 | "already, so I had to buy another one. FYI: The overall \\\n", 104 | "quality has gone done in these types of products, so \\\n", 105 | "they are kind of counting on brand recognition and \\\n", 106 | "consumer loyalty to maintain sales. Got it in about \\\n", 107 | "two days.\n", 108 | "\"\"\"" 109 | ] 110 | }, 111 | { 112 | "cell_type": "code", 113 | "execution_count": null, 114 | "metadata": {}, 115 | "outputs": [], 116 | "source": [ 117 | "prompt = f\"\"\"\n", 118 | "You are a customer service AI assistant.\n", 119 | "Your task is to send an email reply to a valued customer.\n", 120 | "Given the customer email delimited by ```, \\\n", 121 | "Generate a reply to thank the customer for their review.\n", 122 | "If the sentiment is positive or neutral, thank them for \\\n", 123 | "their review.\n", 124 | "If the sentiment is negative, apologize and suggest that \\\n", 125 | "they can reach out to customer service. \n", 126 | "Make sure to use specific details from the review.\n", 127 | "Write in a concise and professional tone.\n", 128 | "Sign the email as `AI customer agent`.\n", 129 | "Customer review: ```{review}```\n", 130 | "Review sentiment: {sentiment}\n", 131 | "\"\"\"\n", 132 | "response = get_completion(prompt)\n", 133 | "print(response)" 134 | ] 135 | }, 136 | { 137 | "cell_type": "markdown", 138 | "metadata": {}, 139 | "source": [] 140 | } 141 | ], 142 | "metadata": { 143 | "kernelspec": { 144 | "display_name": "gpt_index", 145 | "language": "python", 146 | "name": "python3" 147 | }, 148 | "language_info": { 149 | "codemirror_mode": { 150 | "name": "ipython", 151 | "version": 3 152 | }, 153 | "file_extension": ".py", 154 | "mimetype": "text/x-python", 155 | "name": "python", 156 | "nbconvert_exporter": "python", 157 | "pygments_lexer": "ipython3", 158 | "version": "3.8.16" 159 | }, 160 | "orig_nbformat": 4 161 | }, 162 | "nbformat": 4, 163 | "nbformat_minor": 2 164 | } 165 | -------------------------------------------------------------------------------- /chinese_subtitles/prompt_eng_04_summarizing_chinese.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:09,680 3 | 当今世界有如此多的文字，但我们几乎没有人有足够的时间 4 | 5 | 2 6 | 00:00:09,680 --> 00:00:12,760 7 | 来阅读这些文字 8 | 9 | 3 10 | 00:00:12,760 --> 00:00:16,920 11 | 因此，我认为，大语言模型最令人兴奋的应用之一 12 | 13 | 4 14 | 00:00:16,920 --> 00:00:18,660 15 | 是用它来对文本进行总结 16 | 17 | 5 18 | 00:00:18,660 --> 00:00:24,480 19 | 这也是我看到多个团队在不同软件应用中构建的东西 20 | 21 | 6 22 | 00:00:24,480 --> 00:00:26,880 23 | 你可以在ChatGPT网站上这样做 24 | 25 | 7 26 | 00:00:26,880 --> 00:00:31,000 27 | 我经常用这种方式来总结文章 28 | 29 | 8 30 | 00:00:31,000 --> 00:00:33,000 31 | 这样我就可以比以前多读很多文章 32 | 33 | 9 34 | 00:00:33,000 --> 00:00:37,880 35 | 你将在本课程中，学习如何以编程方式来实现这个任务 36 | 37 | 10 38 | 00:00:37,880 --> 00:00:43,520 39 | 因此，让我们深入研究代码，看看你如何使用它来总结文本 40 | 41 | 11 42 | 00:00:43,520 --> 00:00:48,280 43 | 让我们从之前的初始代码开始，导入openai，加载 44 | 45 | 12 46 | 00:00:48,280 --> 00:00:52,620 47 | API Key，这里是get_completion辅助函数 48 | 49 | 13 50 | 00:00:52,620 --> 00:01:00,640 51 | 我将以“总结此产品评价”的任务作为运行示例 52 | 53 | 14 54 | 00:01:00,640 --> 00:01:04,640 55 | 我为女儿的生日买了这只熊猫毛绒玩具，女儿非常喜欢 56 | 57 | 15 58 | 00:01:04,640 --> 00:01:06,320 59 | 无论去哪儿都要带上它，诸如此类 60 | 61 | 16 62 | 00:01:06,320 --> 00:01:13,640 63 | 如果你正在建立一个电子商务网站，并且有大量的评论 64 | 65 | 17 66 | 00:01:13,640 --> 00:01:18,680 67 | 有一个工具可以用来总结冗长的评论，给你提供一种方式 68 | 69 | 18 70 | 00:01:18,680 --> 00:01:24,200 71 | 让你可以浏览更多评论，更好地了解所有客户的想法 72 | 73 | 19 74 | 00:01:24,200 --> 00:01:27,280 75 | 因此，这里有一个生成摘要的提示 76 | 77 | 20 78 | 00:01:27,280 --> 00:01:30,600 79 | 你的任务是对电子商务网站上的产品评论生成一个简短的总结，总结如下 80 | 81 | 21 82 | 00:01:30,600 --> 00:01:36,800 83 | 下面的评论等等，最多只能用30个字 84 | 85 | 22 86 | 00:01:36,800 --> 00:01:42,400 87 | 因此，这个柔软可爱的熊猫毛绒玩具受到女儿的喜爱，但价格有点小贵 88 | 89 | 23 90 | 00:01:42,400 --> 00:01:43,400 91 | 提前到货 92 | 93 | 24 94 | 00:01:43,400 --> 00:01:44,400 95 | 还不错 96 | 97 | 25 98 | 00:01:44,400 --> 00:01:45,400 99 | 这是个相当好的总结 100 | 101 | 26 102 | 00:01:45,400 --> 00:01:49,760 103 | 正如你在前面的视频中看到的，你还可以玩一些东西，比如控制 104 | 105 | 27 106 | 00:01:49,760 --> 00:01:55,720 107 | 字符数或句子数来影响这个摘要的长度 108 | 109 | 28 110 | 00:01:55,720 --> 00:01:59,920 111 | 现在，有时在创建一个摘要时，如果你有一个非常具体的目的 112 | 113 | 29 114 | 00:01:59,920 --> 00:02:05,600 115 | 例如，你想给运输部门提供反馈 116 | 117 | 30 118 | 00:02:05,600 --> 00:02:11,200 119 | 你也可以修改提示来突出这点，这样他们就可以生成一个更好的总结 120 | 121 | 31 122 | 00:02:11,200 --> 00:02:15,600 123 | 适用于你业务中的一个特定群体 124 | 125 | 32 126 | 00:02:15,600 --> 00:02:26,880 127 | 例如，如果我想要给运输部门反馈 128 | 129 | 33 130 | 00:02:26,880 --> 00:02:35,800 131 | 那么我会将关注点集中在商品的运输和交付等方面 132 | 133 | 34 134 | 00:02:35,800 --> 00:02:42,560 135 | 如果我运行这个指令，那么你会得到一个新的摘要 136 | 137 | 35 138 | 00:02:42,560 --> 00:02:49,120 139 | 这次不是以“柔软可爱的熊猫毛绒玩具”开始，而是强调它比预期提前了一天送达 140 | 141 | 36 142 | 00:02:49,120 --> 00:02:53,600 143 | 然后它仍然有其他细节 144 | 145 | 37 146 | 00:02:53,600 --> 00:02:59,640 147 | 或者作为另一个例子，如果我们不是想给运输部门提供反馈 148 | 149 | 38 150 | 00:02:59,640 --> 00:03:09,520 151 | 而是给定价部门提供反馈 152 | 153 | 39 154 | 00:03:09,520 --> 00:03:15,760 155 | 定价部门负责确定产品的价格 156 | 157 | 40 158 | 00:03:15,760 --> 00:03:22,960 159 | 我会告诉它专注于与产品价格和感知价值相关的任何方面 160 | 161 | 41 162 | 00:03:22,960 --> 00:03:25,240 163 | 价值 164 | 165 | 42 166 | 00:03:25,240 --> 00:03:31,520 167 | 那么这将生成一个不同的总结，它说 168 | 169 | 43 170 | 00:03:31,520 --> 00:03:33,480 171 | 也许这个尺寸的价格可能太高了 172 | 173 | 44 174 | 00:03:33,480 --> 00:03:38,640 175 | 现在，在我为运输部门或定价部门生成的摘要中 176 | 177 | 45 178 | 00:03:38,640 --> 00:03:44,400 179 | 它更侧重于与这些特定部门有关的信息 180 | 181 | 46 182 | 00:03:44,400 --> 00:03:50,000 183 | 事实上，你可以随时暂停视频 184 | 185 | 47 186 | 00:03:50,000 --> 00:03:55,360 187 | 然后让它为负责产品客户体验的部门提供有关产品的信息 188 | 189 | 48 190 | 00:03:55,360 --> 00:04:01,920 191 | 或者为你认为对电子商务站点有趣的其他某些方面提供信息 192 | 193 | 49 194 | 00:04:01,920 --> 00:04:07,000 195 | 但在这些摘要中，尽管它生成了与运输有关的信息 196 | 197 | 50 198 | 00:04:07,000 --> 00:04:12,240 199 | 但也有一些其他信息，你可以决定这些信息是否有帮助 200 | 201 | 51 202 | 00:04:12,240 --> 00:04:20,120 203 | 因此，根据你想要总结的方式，你也可以要求它提取信息 204 | 205 | 52 206 | 00:04:20,120 --> 00:04:22,960 207 | 而不是对其进行总结 208 | 209 | 53 210 | 00:04:22,960 --> 00:04:27,920 211 | 因此，这里有一个提示，说你的任务是 212 | 213 | 54 214 | 00:04:27,920 --> 00:04:29,200 215 | 提取相关的信息来反馈给给运输部门 216 | 217 | 55 218 | 00:04:29,200 --> 00:04:35,440 219 | 现在它只是说，产品比预期早到了一天，没有其他的信息 220 | 221 | 56 222 | 00:04:35,440 --> 00:04:41,240 223 | 这在一般的总结中也是有帮助的，但如果想知道运输方面发生了什么 224 | 225 | 57 226 | 00:04:41,240 --> 00:04:45,000 227 | 就不那么具体了 228 | 229 | 58 230 | 00:04:45,000 --> 00:04:50,600 231 | 最后，让我与你分享一个具体的例子，说明如何在工作流程中使用这个方法 232 | 233 | 59 234 | 00:04:50,600 --> 00:04:56,760 235 | 来帮助总结多个评论，使其更容易阅读 236 | 237 | 60 238 | 00:04:56,760 --> 00:04:59,920 239 | 因此，这里有几条评论 240 | 241 | 61 242 | 00:04:59,920 --> 00:05:03,920 243 | 这有点长，但这是对落地灯的第二个评论 244 | 245 | 62 246 | 00:05:03,920 --> 00:05:04,920 247 | 需要一盏灯放在卧室里 248 | 249 | 63 250 | 00:05:04,920 --> 00:05:07,280 251 | 这是关于电动牙刷的第三条评论 252 | 253 | 64 254 | 00:05:07,280 --> 00:05:09,600 255 | 我的牙科保健师推荐的 256 | 257 | 65 258 | 00:05:09,600 --> 00:05:12,640 259 | 关于电动牙刷的评论有点长 260 | 261 | 66 262 | 00:05:12,640 --> 00:05:17,920 263 | 这是关于搅拌机的评论，当时它说是季节性销售的17件套装系统 264 | 265 | 67 266 | 00:05:17,920 --> 00:05:18,920 267 | 诸如此类 268 | 269 | 68 270 | 00:05:18,920 --> 00:05:20,280 271 | 这实际上是很多的文字 272 | 273 | 69 274 | 00:05:20,280 --> 00:05:23,280 275 | 如果你愿意，可以随意暂停视频，读完这些文字 276 | 277 | 70 278 | 00:05:23,280 --> 00:05:28,480 279 | 但是，如果你想知道这些评论者写了什么，而不必停下来阅读所有的 280 | 281 | 71 282 | 00:05:28,480 --> 00:05:29,480 283 | 这方面的细节呢？ 284 | 285 | 72 286 | 00:05:29,480 --> 00:05:41,600 287 | 所以我要把 review_1 设为我们之前展示的那个产品评论 288 | 289 | 73 290 | 00:05:41,600 --> 00:05:45,320 291 | 我将把所有这些评论放到一个列表中 292 | 293 | 74 294 | 00:05:45,320 --> 00:05:53,800 295 | 现在，如果我对这些评论实施一个for循环 296 | 297 | 75 298 | 00:05:53,800 --> 00:06:01,240 299 | 这是我的提示，我要求它最多用20个字来总结它 300 | 301 | 76 302 | 00:06:01,240 --> 00:06:07,320 303 | 然后让它得到响应并打印出来 304 | 305 | 77 306 | 00:06:07,320 --> 00:06:10,760 307 | 让我们运行这个程序 308 | 309 | 78 310 | 00:06:10,760 --> 00:06:16,360 311 | 它打印出的第一条评论是熊猫玩具的评论，灯的总结评论 312 | 313 | 79 314 | 00:06:16,360 --> 00:06:21,880 315 | 牙刷的摘要评论，然后是搅拌器 316 | 317 | 80 318 | 00:06:21,880 --> 00:06:29,800 319 | 因此，如果你有一个网站，你有数百条评论，你可以想象你如何 320 | 321 | 81 322 | 00:06:29,800 --> 00:06:35,360 323 | 使用它来建立一个控制面板，为大量的评论，生成简短的摘要 324 | 325 | 82 326 | 00:06:35,360 --> 00:06:40,760 327 | 这样你或其他人就可以更快地浏览这些评论 328 | 329 | 83 330 | 00:06:40,760 --> 00:06:46,200 331 | 然后，如果他们愿意，可以点击查看原始的长篇评论 332 | 333 | 84 334 | 00:06:46,200 --> 00:06:51,640 335 | 这可以帮助你更高效地了解所有客户的想法 336 | 337 | 85 338 | 00:06:51,640 --> 00:06:55,640 339 | 好了，总结就到这里 340 | 341 | 86 342 | 00:06:55,640 --> 00:07:00,960 343 | 我希望你能想象一下，如果你有任何有许多文本的应用 344 | 345 | 87 346 | 00:07:00,960 --> 00:07:05,680 347 | 你可以使用这样的提示来总结，帮助人们快速了解文字中的内容和各种文本 348 | 349 | 88 350 | 00:07:05,680 --> 00:07:10,240 351 | 可以选择进一步深入挖掘 352 | 353 | 89 354 | 00:07:10,240 --> 00:07:12,520 355 | 如果有需要的话 356 | 357 | 90 358 | 00:07:12,520 --> 00:07:17,640 359 | 在下一个视频中，我们将看到大语言模型的另一个能力，即 360 | 361 | 91 362 | 00:07:17,640 --> 00:07:19,880 363 | 推断的能力 364 | 365 | 92 366 | 00:07:19,880 --> 00:07:24,280 367 | 例如，如果你有一些产品评论数据 368 | 369 | 93 370 | 00:07:24,280 --> 00:07:29,680 371 | 你希望快速了解哪些评论具有积极或消极情感，该怎么办？ 372 | 373 | 94 374 | 00:07:29,680 --> 00:07:49,840 375 | 让我们在下一个视频中看看如何实现这个功能 376 | 377 | -------------------------------------------------------------------------------- /chinese_subtitles/prompt_eng_01_intro_chinese.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:09,000 3 | 欢迎学习《面向开发者的 ChatGPT Prompt 工程》课程 4 | 5 | 2 6 | 00:00:09,000 --> 00:00:14,000 7 | 很高兴能邀请到 Isa Fulford 和我一起授课 8 | 9 | 3 10 | 00:00:14,000 --> 00:00:18,000 11 | 她是 OpenAI 的技术人员之一，曾建立了流行的 12 | 13 | 4 14 | 00:00:18,000 --> 00:00:23,000 15 | ChatGPT 检索插件，她大部分工作是教人们 16 | 17 | 5 18 | 00:00:23,000 --> 00:00:27,000 19 | 如何在产品中使用LLM或大语言模型技术 20 | 21 | 6 22 | 00:00:27,000 --> 00:00:31,000 23 | 她还为OpenAI手册做出了贡献，教会人们如何使用提示 24 | 25 | 7 26 | 00:00:31,000 --> 00:00:32,000 27 | 所以，很高兴有你参与 28 | 29 | 8 30 | 00:00:32,000 --> 00:00:37,000 31 | 我也很高兴能在这里与大家分享一些提示的最佳实践 32 | 33 | 9 34 | 00:00:37,000 --> 00:00:42,000 35 | 在互联网上有很多关于提示资料以及一些文章 36 | 37 | 10 38 | 00:00:42,000 --> 00:00:45,000 39 | 比如，每个人都必须知道的30个提示 40 | 41 | 11 42 | 00:00:45,000 --> 00:00:50,000 43 | 许多人正在使用 ChatGPT 的 Web 用户界面 44 | 45 | 12 46 | 00:00:50,000 --> 00:00:54,000 47 | 来完成特定而且通常是一次性的任务 48 | 49 | 13 50 | 00:00:54,000 --> 00:01:00,000 51 | 但作为一名开发者，我认为LLMs和大语言模型的强大也是不可忽视的 52 | 53 | 14 54 | 00:01:00,000 --> 00:01:04,000 55 | 那就是通过API调用LLM来快速构建软件应用程序 56 | 57 | 15 58 | 00:01:04,000 --> 00:01:08,000 59 | 我认为这仍然被严重低估 60 | 61 | 16 62 | 00:01:08,000 --> 00:01:12,000 63 | 事实上，我在AI Fund的团队，也就是Deep Learning.AI的姐妹公司 64 | 65 | 17 66 | 00:01:12,000 --> 00:01:16,000 67 | 一直在与许多初创公司、不同项目合作 68 | 69 | 18 70 | 00:01:16,000 --> 00:01:18,000 71 | 来应用这些技术 72 | 73 | 19 74 | 00:01:18,000 --> 00:01:23,000 75 | 看到LLM API能够让开发者快速建立一些东西 76 | 77 | 20 78 | 00:01:23,000 --> 00:01:25,000 79 | 这很令人兴奋 80 | 81 | 21 82 | 00:01:25,000 --> 00:01:29,000 83 | 所以在这个课程中，我们将与你分享一些可能性 84 | 85 | 22 86 | 00:01:29,000 --> 00:01:34,000 87 | 你可以做什么，以及如何做的最佳实践 88 | 89 | 23 90 | 00:01:34,000 --> 00:01:36,000 91 | 这里会覆盖很多资料 92 | 93 | 24 94 | 00:01:36,000 --> 00:01:41,000 95 | 首先，你将学习一些软件开发的提示词最佳实践 96 | 97 | 25 98 | 00:01:41,000 --> 00:01:45,000 99 | 然后，我们将涵盖一些常见的用例、总结、推断 100 | 101 | 26 102 | 00:01:45,000 --> 00:01:50,000 103 | 转化、扩展，然后使用LLM建立一个聊天机器人 104 | 105 | 27 106 | 00:01:50,000 --> 00:01:53,000 107 | 我们希望这将激发你 108 | 109 | 28 110 | 00:01:53,000 --> 00:01:55,000 111 | 开发新应用的想象力 112 | 113 | 29 114 | 00:01:55,000 --> 00:01:58,000 115 | 因此，在大型语言模型或LLM的发展中 116 | 117 | 30 118 | 00:01:58,000 --> 00:02:02,000 119 | 大体上有两种类型的LLM，我把它们称为 120 | 121 | 31 122 | 00:02:02,000 --> 00:02:06,000 123 | 基础LLM和指令学习LLM 124 | 125 | 32 126 | 00:02:06,000 --> 00:02:11,000 127 | 基础LLM已经被训练成基于文本训练数据来预测下一个单词 128 | 129 | 33 130 | 00:02:11,000 --> 00:02:15,000 131 | 通常通过互联网和其他来源训练大量数据 132 | 133 | 34 134 | 00:02:15,000 --> 00:02:19,000 135 | 并计算出下一个最可能出现的词是什么 136 | 137 | 35 138 | 00:02:19,000 --> 00:02:24,000 139 | 比如，你输入这个提示："从前有一只独角兽" 140 | 141 | 36 142 | 00:02:24,000 --> 00:02:28,000 143 | 它会进行补全，并预测接下来的几个词是 144 | 145 | 37 146 | 00:02:28,000 --> 00:02:31,000 147 | "和所有独角兽朋友一起生活在一个神奇的森林里" 148 | 149 | 38 150 | 00:02:31,000 --> 00:02:35,000 151 | 但是如果你是用"法国的首都是什么"作为提示 152 | 153 | 39 154 | 00:02:35,000 --> 00:02:40,000 155 | 那么根据互联网上的文章 156 | 157 | 40 158 | 00:02:40,000 --> 00:02:44,000 159 | 很有可能是，基础LLM会以下列方式完成这个任务 160 | 161 | 41 162 | 00:02:44,000 --> 00:02:48,000 163 | 什么是法国最大的城市，什么是法国的人口，等等 164 | 165 | 42 166 | 00:02:48,000 --> 00:02:52,000 167 | 因为互联网上的文章，可能会列出 168 | 169 | 43 170 | 00:02:52,000 --> 00:02:55,000 171 | 关于法国的小测验问题列表 172 | 173 | 44 174 | 00:02:55,000 --> 00:03:00,000 175 | 与此相反，指令学习LLM，发展势头较猛 176 | 177 | 45 178 | 00:03:00,000 --> 00:03:04,000 179 | LLM的研究和实践一直在进行 180 | 181 | 46 182 | 00:03:04,000 --> 00:03:08,000 183 | 一个经过指令学习的LLM已经被训练得能够遵循指令 184 | 185 | 47 186 | 00:03:08,000 --> 00:03:11,000 187 | 因此，如果你问它，法国的首都是什么？ 188 | 189 | 48 190 | 00:03:11,000 --> 00:03:15,000 191 | 它很可能输出法国的首都是巴黎 192 | 193 | 49 194 | 00:03:15,000 --> 00:03:19,000 195 | 因此，指令学习的LLM的典型训练方式是 196 | 197 | 50 198 | 00:03:19,000 --> 00:03:23,000 199 | 从一个在大量文本数据上训练过的基础LLM开始 200 | 201 | 51 202 | 00:03:23,000 --> 00:03:28,000 203 | 然后进一步训练它，用输入和输出来进一步微调它 204 | 205 | 52 206 | 00:03:28,000 --> 00:03:32,000 207 | 这些输入和输出都是指令，也是遵循这些指令的良好尝试 208 | 209 | 53 210 | 00:03:32,000 --> 00:03:36,000 211 | 然后经常使用一种叫做RLHF的技术进一步完善 212 | 213 | 54 214 | 00:03:36,000 --> 00:03:41,000 215 | 从人类反馈中进行强化学习，以使系统能够更好地 216 | 217 | 55 218 | 00:03:41,000 --> 00:03:43,000 219 | 提供帮助并遵循指令 220 | 221 | 56 222 | 00:03:43,000 --> 00:03:47,000 223 | 因为经过指令学习的LLM已经被训练得很有帮助 224 | 225 | 57 226 | 00:03:47,000 --> 00:03:51,000 227 | 诚实且无害，因此，举例来说，它们不太可能输出 228 | 229 | 58 230 | 00:03:51,000 --> 00:03:55,000 231 | 那些与基础LLM相比，会出问题的文本，如有害的输出 232 | 233 | 59 234 | 00:03:55,000 --> 00:03:59,000 235 | 很多实际应用场景已经开始向 236 | 237 | 60 238 | 00:03:59,000 --> 00:04:01,000 239 | 指令学习LLM转移 240 | 241 | 61 242 | 00:04:01,000 --> 00:04:04,000 243 | 你在互联网上找到的一些最佳实践可能更适用于基础LLM 244 | 245 | 62 246 | 00:04:04,000 --> 00:04:08,000 247 | 但对于今天的大多数实际应用情况来说，它们可能不太合适 248 | 249 | 63 250 | 00:04:08,000 --> 00:04:13,000 251 | 我们建议大多数人多关注指令学习LLM 252 | 253 | 64 254 | 00:04:13,000 --> 00:04:17,000 255 | 它们更容易使用，而且由于OpenAI 256 | 257 | 65 258 | 00:04:17,000 --> 00:04:22,000 259 | 和其他LLM公司的工作，也将变得更加安全和一致 260 | 261 | 66 262 | 00:04:22,000 --> 00:04:27,000 263 | 因此，本课程将重点介绍指令学习LLM的最佳实践 264 | 265 | 67 266 | 00:04:27,000 --> 00:04:32,000 267 | 这也是我们建议你在大多数应用中使用的 268 | 269 | 68 270 | 00:04:32,000 --> 00:04:36,000 271 | 在继续之前，我想感谢来自OpenAI 272 | 273 | 69 274 | 00:04:36,000 --> 00:04:39,000 275 | 和DeepLearning.ai的团队 276 | 277 | 70 278 | 00:04:39,000 --> 00:04:42,000 279 | 他们为我和Isa将要介绍的资料做出了贡献 280 | 281 | 71 282 | 00:04:42,000 --> 00:04:45,000 283 | 我非常感谢Andrew Main, Joe Palermo, Boris Power 284 | 285 | 72 286 | 00:04:45,000 --> 00:04:49,000 287 | Ted Sanders，以及来自OpenAI的Lilian Wang 288 | 289 | 73 290 | 00:04:49,000 --> 00:04:53,000 291 | 与我们一起集思广益，审核资料 292 | 293 | 74 294 | 00:04:53,000 --> 00:04:55,000 295 | 并做成这个简短的课程 296 | 297 | 75 298 | 00:04:55,000 --> 00:04:58,000 299 | 我也很感谢DeepLearning方面的工作 300 | 301 | 76 302 | 00:04:58,000 --> 00:05:01,000 303 | Jeff Ludwig、Eddie Hsu和Tommy Nelson的工作 304 | 305 | 77 306 | 00:05:01,000 --> 00:05:06,000 307 | 因此，当你使用一个指令学习LLM时，可以将其看作是 308 | 309 | 78 310 | 00:05:06,000 --> 00:05:10,000 311 | 向另一个人发出指令，比如一个聪明但不知道 312 | 313 | 79 314 | 00:05:10,000 --> 00:05:12,000 315 | 任务细节的人 316 | 317 | 80 318 | 00:05:12,000 --> 00:05:16,000 319 | 因此，当一个LLM不工作时，有时是因为指令 320 | 321 | 81 322 | 00:05:16,000 --> 00:05:17,000 323 | 不够清楚 324 | 325 | 82 326 | 00:05:17,000 --> 00:05:20,000 327 | 例如，如果你说，请给我写一些 328 | 329 | 83 330 | 00:05:20,000 --> 00:05:22,000 331 | 关于艾伦-图灵的东西 332 | 333 | 84 334 | 00:05:22,000 --> 00:05:26,000 335 | 那么，除此之外，明确以下几点也会有所帮助 336 | 337 | 85 338 | 00:05:26,000 --> 00:05:30,000 339 | 你想让文章侧重于他的科学工作 340 | 341 | 86 342 | 00:05:30,000 --> 00:05:34,000 343 | 他的个人生活、他在历史上的作用或其他方向 344 | 345 | 87 346 | 00:05:34,000 --> 00:05:39,000 347 | 如果你指定了想要文本的语调 348 | 349 | 88 350 | 00:05:39,000 --> 00:05:43,000 351 | 它是否会根据指定的语调来写，比如像专业记者写的那样 352 | 353 | 89 354 | 00:05:43,000 --> 00:05:46,000 355 | 还是更像随手写给朋友的轻松语气？ 356 | 357 | 90 358 | 00:05:46,000 --> 00:05:47,000 359 | 这一点是成立的 360 | 361 | 91 362 | 00:05:47,000 --> 00:05:49,000 363 | LLM会产生你想要的东西 364 | 365 | 92 366 | 00:05:49,000 --> 00:05:52,000 367 | 当然，如果你想象是自己在问，比如 368 | 369 | 93 370 | 00:05:52,000 --> 00:05:56,000 371 | 一个刚毕业的大学生为你完成这项任务 372 | 373 | 94 374 | 00:05:56,000 --> 00:05:59,000 375 | 如果你甚至可以指定他们应该阅读哪些文本片段 376 | 377 | 95 378 | 00:05:59,000 --> 00:06:02,000 379 | 提前写出这篇关于艾伦-图灵的文章 380 | 381 | 96 382 | 00:06:02,000 --> 00:06:06,000 383 | 那么，这能成功让那个刚毕业的大学生 384 | 385 | 97 386 | 00:06:06,000 --> 00:06:09,000 387 | 更好的为你完成这项任务 388 | 389 | 98 390 | 00:06:09,000 --> 00:06:13,000 391 | 因此，在接下来的视频中，你会看到如何清晰且明确 392 | 393 | 99 394 | 00:06:13,000 --> 00:06:17,000 395 | 描述提示，这是一个重要的LLM提示准则 396 | 397 | 100 398 | 00:06:17,000 --> 00:06:21,000 399 | 而且你还将从Isa那里学到了第二个提示词准则 400 | 401 | 101 402 | 00:06:21,000 --> 00:06:24,000 403 | 那就是给LLM时间来思考 404 | 405 | 102 406 | 00:06:24,000 --> 00:06:29,000 407 | 因此，让我们继续观看下一个视频 408 | 409 | -------------------------------------------------------------------------------- /subtitles/styled/prompt_eng_09_conclusion.ass: -------------------------------------------------------------------------------- 1 | [Script Info] 2 | ; Script generated by Aegisub 3.3.3 3 | ; http://www.aegisub.org/ 4 | Title: Default Aegisub file 5 | ScriptType: v4.00+ 6 | WrapStyle: 0 7 | ScaledBorderAndShadow: yes 8 | YCbCr Matrix: TV.709 9 | PlayResX: 1920 10 | PlayResY: 1080 11 | 12 | [Aegisub Project Garbage] 13 | Active Line: 5 14 | 15 | [V4+ Styles] 16 | Format: Name, Fontname, Fontsize, PrimaryColour, SecondaryColour, OutlineColour, BackColour, Bold, Italic, Underline, StrikeOut, ScaleX, ScaleY, Spacing, Angle, BorderStyle, Outline, Shadow, Alignment, MarginL, MarginR, MarginV, Encoding 17 | Style: Default,Arial,48,&H00FFFFFF,&H000000FF,&H00000000,&H00000000,0,0,0,0,100,100,0,0,1,2,2,2,10,10,10,1 18 | Style: 蓝中字,FZHei-B01,60,&H00FFFFFF,&H000000FF,&H0ABD631D,&H64000000,0,0,0,0,100,100,0,0,1,3,1,2,10,10,10,1 19 | Style: 灰英字,Helvetica Neue,40,&H00FFFFFF,&H000000FF,&H0A3C3C3C,&H64000000,0,0,0,0,100,100,0,0,1,3,1,2,10,10,10,1 20 | 21 | [Events] 22 | Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text 23 | Dialogue: 0,0:00:05.00,0:00:08.84,灰英字,,0,0,0,,Congratulations on making it to the end of this short course 24 | Dialogue: 0,0:00:08.84,0:00:11.28,灰英字,,0,0,0,,In summary, in this short course 25 | Dialogue: 0,0:00:11.28,0:00:14.38,灰英字,,0,0,0,,you've learned about two key principles for prompting 26 | Dialogue: 0,0:00:14.38,0:00:17.48,灰英字,,0,0,0,,right clearance for specific instructions and when it's appropriate 27 | Dialogue: 0,0:00:17.48,0:00:19.16,灰英字,,0,0,0,,give the model time to think 28 | Dialogue: 0,0:00:19.16,0:00:22.08,灰英字,,0,0,0,,You also learned about iterative prompt development and how 29 | Dialogue: 0,0:00:22.08,0:00:26.20,灰英字,,0,0,0,,having a process to get to the prompt is right for your application is key 30 | Dialogue: 0,0:00:26.20,0:00:28.80,灰英字,,0,0,0,,And we went through a few capabilities of 31 | Dialogue: 0,0:00:28.80,0:00:31.64,灰英字,,0,0,0,,large language models that are useful for many applications 32 | Dialogue: 0,0:00:31.64,0:00:36.32,灰英字,,0,0,0,,specifically summarizing, inferring, transforming, and expanding 33 | Dialogue: 0,0:00:36.32,0:00:39.88,灰英字,,0,0,0,,And you also saw how to build a custom chatbot 34 | Dialogue: 0,0:00:39.88,0:00:43.32,灰英字,,0,0,0,,That was a lot that you learned in just one short course 35 | Dialogue: 0,0:00:43.32,0:00:46.24,灰英字,,0,0,0,,and I hope you enjoyed going through these materials 36 | Dialogue: 0,0:00:46.24,0:00:47.92,灰英字,,0,0,0,,We hope you'll come up with some ideas for 37 | Dialogue: 0,0:00:47.92,0:00:50.12,灰英字,,0,0,0,,applications that you can build yourself now 38 | Dialogue: 0,0:00:50.12,0:00:53.12,灰英字,,0,0,0,,And please go try this out and let us know what you come up with 39 | Dialogue: 0,0:00:53.12,0:00:54.72,灰英字,,0,0,0,,No application is too small 40 | Dialogue: 0,0:00:54.72,0:00:57.00,灰英字,,0,0,0,,It's fine to start with something that's kind of 41 | Dialogue: 0,0:00:57.00,0:00:59.80,灰英字,,0,0,0,,a very small project with maybe a little bit of utility 42 | Dialogue: 0,0:00:59.80,0:01:01.80,灰英字,,0,0,0,,or maybe it's not even useful at all 43 | Dialogue: 0,0:01:01.80,0:01:03.32,灰英字,,0,0,0,,It's just something fun 44 | Dialogue: 0,0:01:03.32,0:01:06.20,灰英字,,0,0,0,,Yeah. And I find playing with these models actually really fun 45 | Dialogue: 0,0:01:06.20,0:01:07.44,灰英字,,0,0,0,,So go play with it 46 | Dialogue: 0,0:01:07.44,0:01:08.64,灰英字,,0,0,0,,Yeah, I agree. It's a good 47 | Dialogue: 0,0:01:08.64,0:01:11.08,灰英字,,0,0,0,,a good weekend activity, speaking from experience 48 | Dialogue: 0,0:01:11.08,0:01:14.96,灰英字,,0,0,0,,And just, you know 49 | Dialogue: 0,0:01:14.96,0:01:18.52,灰英字,,0,0,0,,please use the learnings from your first project to build a better second project 50 | Dialogue: 0,0:01:18.52,0:01:21.36,灰英字,,0,0,0,,and you know, maybe even a better third project, so on 51 | Dialogue: 0,0:01:21.36,0:01:26.96,灰英字,,0,0,0,,That's kind of how I have kind of grown over time using these models myself as well 52 | Dialogue: 0,0:01:26.96,0:01:31.36,灰英字,,0,0,0,,Or if you have an idea for a bigger project already, just, just go for it 53 | Dialogue: 0,0:01:31.36,0:01:33.96,灰英字,,0,0,0,,And, you know, as a reminder 54 | Dialogue: 0,0:01:33.96,0:01:37.20,灰英字,,0,0,0,,these kind of large language models are a very powerful technology 55 | Dialogue: 0,0:01:37.20,0:01:41.12,灰英字,,0,0,0,,So it kind of goes without saying that we ask you to use them responsibly 56 | Dialogue: 0,0:01:41.12,0:01:43.96,灰英字,,0,0,0,,and please only build things that will have a positive impact 57 | Dialogue: 0,0:01:43.96,0:01:46.24,灰英字,,0,0,0,,Yeah, fully agree. I think in this age 58 | Dialogue: 0,0:01:46.24,0:01:49.48,灰英字,,0,0,0,,people that build AI systems can have a huge impact on others 59 | Dialogue: 0,0:01:49.48,0:01:54.96,灰英字,,0,0,0,,So it's more important than ever that all of us only use these tools responsibly 60 | Dialogue: 0,0:01:54.96,0:01:59.72,灰英字,,0,0,0,,And I think building large language model-based applications 61 | Dialogue: 0,0:01:59.72,0:02:02.88,灰英字,,0,0,0,,is just a very exciting and growing field right now 62 | Dialogue: 0,0:02:02.88,0:02:05.16,灰英字,,0,0,0,,And now that you've finished this course 63 | Dialogue: 0,0:02:05.16,0:02:08.24,灰英字,,0,0,0,,I think you now have a wealth of knowledge to let 64 | Dialogue: 0,0:02:08.24,0:02:11.32,灰英字,,0,0,0,,you build things that few people today know how to 65 | Dialogue: 0,0:02:11.32,0:02:17.28,灰英字,,0,0,0,,So I hope you also help us to spread the word and encourage others to take this course too 66 | Dialogue: 0,0:02:17.28,0:02:20.52,灰英字,,0,0,0,,In closing, I hope you had fun doing this course 67 | Dialogue: 0,0:02:20.52,0:02:23.28,灰英字,,0,0,0,,and I want to thank you for finishing this course 68 | Dialogue: 0,0:02:23.28,0:02:27.12,灰英字,,0,0,0,,and both Isa and I look forward to hearing about the amazing things that you've built 69 | Dialogue: 0,0:00:05.00,0:00:08.84,蓝中字,,0,0,0,,祝贺你在这个短期课程中走到了最后 70 | Dialogue: 0,0:00:08.84,0:00:11.28,蓝中字,,0,0,0,,综上所述，在这个短课程中 71 | Dialogue: 0,0:00:11.28,0:00:14.38,蓝中字,,0,0,0,,你已经了解了提示的两个关键原则 72 | Dialogue: 0,0:00:14.38,0:00:17.48,蓝中字,,0,0,0,,正确且清晰的描述具体指示，以及知道什么时候使用 73 | Dialogue: 0,0:00:17.48,0:00:19.16,蓝中字,,0,0,0,,给予模型思考的时间 74 | Dialogue: 0,0:00:19.16,0:00:22.08,蓝中字,,0,0,0,,你还了解了迭代提示的发展以及如何 75 | Dialogue: 0,0:00:22.08,0:00:26.20,蓝中字,,0,0,0,,有一个过程来获得适合你的应用的提示是关键 76 | Dialogue: 0,0:00:26.20,0:00:28.80,蓝中字,,0,0,0,,我们还学习了一些大语言模型的功能 77 | Dialogue: 0,0:00:28.80,0:00:31.64,蓝中字,,0,0,0,,这些功能对许多应用是有用的 78 | Dialogue: 0,0:00:31.64,0:00:36.32,蓝中字,,0,0,0,,特别是总结、推理、转换和扩展 79 | Dialogue: 0,0:00:36.32,0:00:39.88,蓝中字,,0,0,0,,你还看到了如何建立一个定制的聊天机器人 80 | Dialogue: 0,0:00:39.88,0:00:43.32,蓝中字,,0,0,0,,在短短的课程中，你学到了很多东西 81 | Dialogue: 0,0:00:43.32,0:00:46.24,蓝中字,,0,0,0,,我希望你喜欢阅读这些材料 82 | Dialogue: 0,0:00:46.24,0:00:47.92,蓝中字,,0,0,0,,我们希望你能想出一些想法 83 | Dialogue: 0,0:00:47.92,0:00:50.12,蓝中字,,0,0,0,,我们希望你能想出一些可以自己构建的应用 84 | Dialogue: 0,0:00:50.12,0:00:53.12,蓝中字,,0,0,0,,并请你去尝试一下，让我们知道你想出了什么 85 | Dialogue: 0,0:00:53.12,0:00:54.72,蓝中字,,0,0,0,,任何应用都不小 86 | Dialogue: 0,0:00:54.72,0:00:57.00,蓝中字,,0,0,0,,从一些很小的项目开始也是可以的 87 | Dialogue: 0,0:00:57.00,0:00:59.80,蓝中字,,0,0,0,,一个非常小的项目，也许实用 88 | Dialogue: 0,0:00:59.80,0:01:01.80,蓝中字,,0,0,0,,也可能毫无用处 89 | Dialogue: 0,0:01:01.80,0:01:03.32,蓝中字,,0,0,0,,这只是一些有趣的东西 90 | Dialogue: 0,0:01:03.32,0:01:06.20,蓝中字,,0,0,0,,是的。我发现玩这些模型非常有趣 91 | Dialogue: 0,0:01:06.20,0:01:07.44,蓝中字,,0,0,0,,所以放心去玩吧 92 | Dialogue: 0,0:01:07.44,0:01:08.64,蓝中字,,0,0,0,,是的，我同意。从我的经验来看 93 | Dialogue: 0,0:01:08.64,0:01:11.08,蓝中字,,0,0,0,,这是一个很好的周末活动 94 | Dialogue: 0,0:01:11.08,0:01:14.96,蓝中字,,0,0,0,,而且，你可以通过第一个项目获得的经验教训 95 | Dialogue: 0,0:01:14.96,0:01:18.52,蓝中字,,0,0,0,,来构建第二个更好的项目 96 | Dialogue: 0,0:01:18.52,0:01:21.36,蓝中字,,0,0,0,,甚至可能是更好的第三个项目，等等 97 | Dialogue: 0,0:01:21.36,0:01:26.96,蓝中字,,0,0,0,,这就是我自己使用这些模型逐渐成长的方式 98 | Dialogue: 0,0:01:26.96,0:01:31.36,蓝中字,,0,0,0,,或者如果你已经有了一个更大的项目想法，就去做吧 99 | Dialogue: 0,0:01:31.36,0:01:33.96,蓝中字,,0,0,0,,作为一个提醒 100 | Dialogue: 0,0:01:33.96,0:01:37.20,蓝中字,,0,0,0,,这些大语言模型是一种非常强大的技术 101 | Dialogue: 0,0:01:37.20,0:01:41.12,蓝中字,,0,0,0,,所以不言而喻，我们要求你负责任地使用它们 102 | Dialogue: 0,0:01:41.12,0:01:43.96,蓝中字,,0,0,0,,请只建立会产生积极影响的东西 103 | Dialogue: 0,0:01:43.96,0:01:46.24,蓝中字,,0,0,0,,是的，完全同意。我认为在这个时代 104 | Dialogue: 0,0:01:46.24,0:01:49.48,蓝中字,,0,0,0,,建立AI系统的人可以对其他人产生巨大的影响 105 | Dialogue: 0,0:01:49.48,0:01:54.96,蓝中字,,0,0,0,,因此，我们所有人都要负责任地使用这些工具，这一点比以往任何时候都重要 106 | Dialogue: 0,0:01:54.96,0:01:59.72,蓝中字,,0,0,0,,我认为建立基于大语言模型的应用程序 107 | Dialogue: 0,0:01:59.72,0:02:02.88,蓝中字,,0,0,0,,是一个非常令人兴奋和不断增长的领域 108 | Dialogue: 0,0:02:02.88,0:02:05.16,蓝中字,,0,0,0,,现在你已经完成了这个课程 109 | Dialogue: 0,0:02:05.16,0:02:08.24,蓝中字,,0,0,0,,我想你现在已经有了丰富的知识 110 | Dialogue: 0,0:02:08.24,0:02:11.32,蓝中字,,0,0,0,,可以让你构建少数人知道如何构建的东西 111 | Dialogue: 0,0:02:11.32,0:02:17.28,蓝中字,,0,0,0,,与此同时，我希望你也能帮助我们传播信息，鼓励其他人也参加这个课程 112 | Dialogue: 0,0:02:17.28,0:02:20.52,蓝中字,,0,0,0,,最后，我希望你在做这个课程时很开心 113 | Dialogue: 0,0:02:20.52,0:02:23.28,蓝中字,,0,0,0,,同时也感谢你能完成这个课程 114 | Dialogue: 0,0:02:23.28,0:02:27.12,蓝中字,,0,0,0,,Isa和我都期待着听到你所建立的惊人的东西 -------------------------------------------------------------------------------- /chinese_subtitles/prompt_eng_07_expanding_chinese.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:08,200 3 | 扩写是指将一篇较短的文本进行扩写的工作 4 | 5 | 2 6 | 00:00:08,200 --> 00:00:11,140 7 | 如一套指示或一个主题列表 8 | 9 | 3 10 | 00:00:11,140 --> 00:00:14,900 11 | 并让大语言模型生成一个较长的文本 12 | 13 | 4 14 | 00:00:14,900 --> 00:00:18,260 15 | 例如一封电子邮件或一篇关于某些主题的文章 16 | 17 | 5 18 | 00:00:18,260 --> 00:00:20,100 19 | 这方面有一些很好的用途 20 | 21 | 6 22 | 00:00:20,100 --> 00:00:23,960 23 | 例如，如果你使用一个大语言模型作为头脑风暴的伙伴 24 | 25 | 7 26 | 00:00:23,960 --> 00:00:26,300 27 | 但是，我也想承认，有一些 28 | 29 | 8 30 | 00:00:26,300 --> 00:00:28,500 31 | 有问题的使用案例 32 | 33 | 9 34 | 00:00:28,500 --> 00:00:32,500 35 | 例如，如果有人用它来产生大量的垃圾邮件 36 | 37 | 10 38 | 00:00:32,500 --> 00:00:36,100 39 | 因此，当你使用大语言模型的这些能力时 40 | 41 | 11 42 | 00:00:36,100 --> 00:00:41,340 43 | 请以负责任的方式和帮助人们的方式来使用它 44 | 45 | 12 46 | 00:00:41,340 --> 00:00:45,280 47 | 在这个视频中，我们将通过一个例子来说明 48 | 49 | 13 50 | 00:00:45,280 --> 00:00:49,900 51 | 如何使用语言模型，生成一个基于某些信息的个性化电子邮件 52 | 53 | 14 54 | 00:00:49,900 --> 00:00:53,260 55 | 这封邮件自称是来自一个AI机器人 56 | 57 | 15 58 | 00:00:53,260 --> 00:00:56,140 59 | 正如恩达提到的，这一点非常重要 60 | 61 | 16 62 | 00:00:56,140 --> 00:00:58,500 63 | 我们还将使用另一个 64 | 65 | 17 66 | 00:00:58,500 --> 00:01:01,260 67 | 该模型的另一个输入参数，叫做temperature（温度） 68 | 69 | 18 70 | 00:01:01,260 --> 00:01:04,780 71 | 这允许你改变模型的探索和多样性的程度 72 | 73 | 19 74 | 00:01:04,780 --> 00:01:08,760 75 | 这允许你改变模型响应的探索和多样性 76 | 77 | 20 78 | 00:01:08,760 --> 00:01:10,980 79 | 因此，让我们开始吧 80 | 81 | 21 82 | 00:01:10,980 --> 00:01:12,740 83 | 在我们开始之前 84 | 85 | 22 86 | 00:01:12,740 --> 00:01:15,780 87 | 我们要做一些常规的设置 88 | 89 | 23 90 | 00:01:15,780 --> 00:01:18,400 91 | 所以要设置OpenAI Python软件包 92 | 93 | 24 94 | 00:01:18,400 --> 00:01:22,060 95 | 然后定义我们的辅助函数，get_completion 96 | 97 | 25 98 | 00:01:22,060 --> 00:01:26,460 99 | 现在我们要为客户的评论写一个自定义的电子邮件回复 100 | 101 | 26 102 | 00:01:26,460 --> 00:01:29,940 103 | 所以给定一个客户评论和情绪 104 | 105 | 27 106 | 00:01:29,940 --> 00:01:32,680 107 | 我们将生成一个自定义的回复 108 | 109 | 28 110 | 00:01:32,680 --> 00:01:37,100 111 | 现在我们要使用语言模型来生成 112 | 113 | 29 114 | 00:01:37,100 --> 00:01:41,900 115 | 根据客户的评论，给客户发送一封定制的电子邮件 116 | 117 | 30 118 | 00:01:41,900 --> 00:01:44,020 119 | 和评论的情绪 120 | 121 | 31 122 | 00:01:44,020 --> 00:01:47,100 123 | 所以我们已经提取了情绪 124 | 125 | 32 126 | 00:01:47,100 --> 00:01:52,740 127 | 使用我们在推理视频中看到的那种提示 128 | 129 | 33 130 | 00:01:52,740 --> 00:01:57,020 131 | 然后这就是客户对搅拌机的评论 132 | 133 | 34 134 | 00:01:57,100 --> 00:02:02,740 135 | 现在我们要根据情感来定制回复 136 | 137 | 35 138 | 00:02:02,740 --> 00:02:05,180 139 | 所以这里的指令是 140 | 141 | 36 142 | 00:02:05,180 --> 00:02:08,260 143 | 你是一个客户服务的AI助理 144 | 145 | 37 146 | 00:02:08,260 --> 00:02:11,180 147 | 你的任务是给一个有价值的客户发送电子邮件回复 148 | 149 | 38 150 | 00:02:11,180 --> 00:02:14,540 151 | 给出客户的电子邮件，以三个`符号分割 152 | 153 | 39 154 | 00:02:14,540 --> 00:02:16,860 155 | 产生一个回复，以感谢客户的评论 156 | 157 | 40 158 | 00:02:16,860 --> 00:02:18,780 159 | 如果情绪是积极的或中性的 160 | 161 | 41 162 | 00:02:18,780 --> 00:02:20,460 163 | 感谢他们的评论 164 | 165 | 42 166 | 00:02:20,460 --> 00:02:21,980 167 | 如果情绪是负面的 168 | 169 | 43 170 | 00:02:21,980 --> 00:02:25,260 171 | 道歉，并建议他们可以联系客户服务部 172 | 173 | 44 174 | 00:02:25,260 --> 00:02:27,820 175 | 确保使用评论中的具体细节 176 | 177 | 45 178 | 00:02:27,820 --> 00:02:29,820 179 | 用简明和专业的语气来写 180 | 181 | 46 182 | 00:02:29,820 --> 00:02:32,460 183 | 并以AI客户代理的身份在邮件中签名 184 | 185 | 47 186 | 00:02:32,460 --> 00:02:34,820 187 | 当你使用一个语言模型 188 | 189 | 48 190 | 00:02:34,820 --> 00:02:37,500 191 | 来生成你要展示给用户的文本 192 | 193 | 49 194 | 00:02:37,500 --> 00:02:40,860 195 | 有这样的透明度是非常重要的 196 | 197 | 50 198 | 00:02:40,860 --> 00:02:45,700 199 | 并让用户知道他们看到的文本是由AI生成的 200 | 201 | 51 202 | 00:02:45,700 --> 00:02:49,780 203 | 然后我们就输入客户的评论和评论的情绪 204 | 205 | 52 206 | 00:02:49,780 --> 00:02:53,220 207 | 还要注意，这部分不一定很重要 208 | 209 | 53 210 | 00:02:53,220 --> 00:02:57,180 211 | 因为我们实际上也可以使用这个提示来提取评论的情绪 212 | 213 | 54 214 | 00:02:57,180 --> 00:02:59,300 215 | 然后在后续步骤中，编写电子邮件 216 | 217 | 55 218 | 00:02:59,300 --> 00:03:01,140 219 | 但是，只是为了这个例子的目的 220 | 221 | 56 222 | 00:03:01,140 --> 00:03:04,300 223 | 好吧，我们已经从评论中提取了情感 224 | 225 | 57 226 | 00:03:04,300 --> 00:03:08,500 227 | 所以在这里我们有一个给客户的回复 228 | 229 | 58 230 | 00:03:08,500 --> 00:03:13,580 231 | 它涉及客户在评论中提到的细节 232 | 233 | 59 234 | 00:03:13,580 --> 00:03:16,620 235 | 并且按照我们的指示 236 | 237 | 60 238 | 00:03:16,620 --> 00:03:19,100 239 | 建议他们与客户服务部联系 240 | 241 | 61 242 | 00:03:19,100 --> 00:03:23,100 243 | 因为这只是一个AI客户服务代理 244 | 245 | 62 246 | 00:03:23,740 --> 00:03:30,420 247 | 接下来，我们将使用语言模型的一个参数，叫做temperature 248 | 249 | 63 250 | 00:03:30,420 --> 00:03:36,460 251 | 这将使我们能够改变模型响应的多样性 252 | 253 | 64 254 | 00:03:36,460 --> 00:03:38,900 255 | 所以你可以把temperature看成 256 | 257 | 65 258 | 00:03:38,900 --> 00:03:43,380 259 | 模型的探索或随机性的程度 260 | 261 | 66 262 | 00:03:43,380 --> 00:03:46,060 263 | 因此，对于这个特定的短语 264 | 265 | 67 266 | 00:03:46,060 --> 00:03:47,500 267 | 我最喜欢的食物是 268 | 269 | 68 270 | 00:03:47,500 --> 00:03:51,820 271 | 模型预测的最有可能的下一个词是披萨 272 | 273 | 69 274 | 00:03:51,820 --> 00:03:53,580 275 | 而接下来的两个最可能的 276 | 277 | 70 278 | 00:03:53,580 --> 00:03:55,900 279 | 它建议是寿司和玉米饼 280 | 281 | 71 282 | 00:03:55,900 --> 00:03:57,660 283 | 因此，在temperature为零的情况下 284 | 285 | 72 286 | 00:03:57,660 --> 00:04:00,060 287 | 模型总是会选择最可能的下一个词 288 | 289 | 73 290 | 00:04:00,060 --> 00:04:01,460 291 | 在这个例子中是比萨 292 | 293 | 74 294 | 00:04:01,460 --> 00:04:03,260 295 | 而在更高的temperature下 296 | 297 | 75 298 | 00:04:03,260 --> 00:04:07,100 299 | 它也会选择一个不太可能的词 300 | 301 | 76 302 | 00:04:07,100 --> 00:04:09,340 303 | 甚至在更高的temperature下 304 | 305 | 77 306 | 00:04:09,340 --> 00:04:10,700 307 | 它有可能选择玉米饼 308 | 309 | 78 310 | 00:04:10,700 --> 00:04:15,380 311 | 虽然这只有5%的机率被选中 312 | 313 | 79 314 | 00:04:15,380 --> 00:04:21,660 315 | 你可以想象，当模型继续这个最后的响应 316 | 317 | 80 318 | 00:04:21,660 --> 00:04:23,300 319 | 所以我最喜欢的食物是比萨饼 320 | 321 | 81 322 | 00:04:23,300 --> 00:04:25,420 323 | 而且它继续产生更多的词 324 | 325 | 82 326 | 00:04:25,420 --> 00:04:30,580 327 | 这个响应会与第一个响应发生分歧 328 | 329 | 83 330 | 00:04:30,580 --> 00:04:32,340 331 | 也就是我最喜欢的食物是玉米饼 332 | 333 | 84 334 | 00:04:32,340 --> 00:04:34,100 335 | 所以随着模型的继续 336 | 337 | 85 338 | 00:04:34,100 --> 00:04:37,060 339 | 这两个响应将变得越来越不同 340 | 341 | 86 342 | 00:04:37,060 --> 00:04:40,500 343 | 一般来说，在构建你想要的应用程序时 344 | 345 | 87 346 | 00:04:40,500 --> 00:04:43,420 347 | 一个可预测的响应 348 | 349 | 88 350 | 00:04:43,420 --> 00:04:45,420 351 | 我建议使用temperature=0 352 | 353 | 89 354 | 00:04:45,420 --> 00:04:46,780 355 | 在所有这些视频中 356 | 357 | 90 358 | 00:04:46,780 --> 00:04:48,900 359 | 我们一直在使用temperature=0 360 | 361 | 91 362 | 00:04:48,900 --> 00:04:51,380 363 | 我认为，如果你试图建立一个系统 364 | 365 | 92 366 | 00:04:51,380 --> 00:04:54,660 367 | 可靠和可预测的系统，你应该使用这个 368 | 369 | 93 370 | 00:04:54,660 --> 00:04:58,780 371 | 如果你想以更有创意的方式使用这个模型 372 | 373 | 94 374 | 00:04:58,780 --> 00:05:04,220 375 | 你可能希望有更多不同的输出 376 | 377 | 95 378 | 00:05:04,220 --> 00:05:06,860 379 | 你可能想用更高的temperature 380 | 381 | 96 382 | 00:05:06,860 --> 00:05:11,620 383 | 那么，现在让我们把刚才的提示也拿出来 384 | 385 | 97 386 | 00:05:11,620 --> 00:05:14,220 387 | 让我们试着生成一封电子邮件 388 | 389 | 98 390 | 00:05:14,220 --> 00:05:16,140 391 | 但让我们使用一个更高的temperature 392 | 393 | 99 394 | 00:05:16,140 --> 00:05:21,540 395 | 所以在我们的get_completion函数中，我们在整个视频中一直在使用 396 | 397 | 100 398 | 00:05:21,540 --> 00:05:25,100 399 | 我们已经指定了一个模型和一个temperature 400 | 401 | 101 402 | 00:05:25,100 --> 00:05:26,540 403 | 但我们已经把它们设置为默认值 404 | 405 | 102 406 | 00:05:26,540 --> 00:05:29,420 407 | 所以现在让我们试着改变temperature 408 | 409 | 103 410 | 00:05:30,140 --> 00:05:32,860 411 | 所以我们使用提示 412 | 413 | 104 414 | 00:05:32,860 --> 00:05:39,340 415 | 然后让我们试试temperature=0.7 416 | 417 | 105 418 | 00:05:43,620 --> 00:05:45,340 419 | 于是用temperature=0 420 | 421 | 106 422 | 00:05:45,340 --> 00:05:47,220 423 | 每次你执行同样的提示时 424 | 425 | 107 426 | 00:05:47,220 --> 00:05:50,580 427 | 你应该期待同样的完成 428 | 429 | 108 430 | 00:05:50,580 --> 00:05:52,580 431 | 而在temperature为0.7时 432 | 433 | 109 434 | 00:05:52,580 --> 00:05:55,060 435 | 你每次都会得到一个不同的输出 436 | 437 | 110 438 | 00:05:55,060 --> 00:05:58,100 439 | 所以这里我们有我们的电子邮件 440 | 441 | 111 442 | 00:05:58,100 --> 00:06:00,980 443 | 正如你所看到的，它与我们之前收到的电子邮件 444 | 445 | 112 446 | 00:06:00,980 --> 00:06:02,900 447 | 不同 448 | 449 | 113 450 | 00:06:02,900 --> 00:06:08,060 451 | 让我们再执行一次，以显示我们将再次得到一个不同的电子邮件 452 | 453 | 114 454 | 00:06:08,060 --> 00:06:11,340 455 | 这里我们有另一封不同的电子邮件 456 | 457 | 115 458 | 00:06:11,340 --> 00:06:16,780 459 | 所以我建议你自己玩一下temperature 460 | 461 | 116 462 | 00:06:16,780 --> 00:06:20,820 463 | 也许你现在可以暂停视频，并尝试这个提示 464 | 465 | 117 466 | 00:06:20,820 --> 00:06:22,620 467 | 用各种不同的temperature 468 | 469 | 118 470 | 00:06:22,620 --> 00:06:25,100 471 | 只是为了看看输出是如何变化的 472 | 473 | 119 474 | 00:06:25,100 --> 00:06:28,500 475 | 因此，总结一下，在较高的temperature下 476 | 477 | 120 478 | 00:06:28,500 --> 00:06:31,700 479 | 模型的输出更加随机 480 | 481 | 121 482 | 00:06:31,700 --> 00:06:34,500 483 | 你几乎可以认为，在较高temperature下 484 | 485 | 122 486 | 00:06:34,500 --> 00:06:36,500 487 | 助理更容易分心 488 | 489 | 123 490 | 00:06:36,500 --> 00:06:38,220 491 | 但也许更有创造力 492 | 493 | 124 494 | 00:06:39,620 --> 00:06:42,260 495 | 在下一个视频中，我们将进一步讨论 496 | 497 | 125 498 | 00:06:42,260 --> 00:06:44,540 499 | 对话补全终端格式 500 | 501 | 126 502 | 00:06:44,540 --> 00:07:00,060 503 | 以及你如何使用这种格式创建一个自定义的聊天机器人 504 | 505 | -------------------------------------------------------------------------------- /course_notebooks/l8-chatbot.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "attachments": {}, 5 | "cell_type": "markdown", 6 | "metadata": {}, 7 | "source": [ 8 | "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/GitHubDaily/ChatGPT-Prompt-Engineering-for-Developers-in-Chinese/blob/master/assets/8-Chatbot/l8-chatbot.ipynb)\n", 9 | "# The Chat Format\n", 10 | "\n", 11 | "In this notebook, you will explore how you can utilize the chat format to have extended conversations with chatbots personalized or specialized for specific tasks or behaviors.\n", 12 | "\n", 13 | "## Setup" 14 | ] 15 | }, 16 | { 17 | "cell_type": "code", 18 | "execution_count": null, 19 | "metadata": {}, 20 | "outputs": [], 21 | "source": [ 22 | "# Basic congfig \n", 23 | "# Install basic package and set key\n", 24 | "!pip install openai\n", 25 | "!export OPENAI_API_KEY='sk-...'" 26 | ] 27 | }, 28 | { 29 | "cell_type": "code", 30 | "execution_count": null, 31 | "metadata": {}, 32 | "outputs": [], 33 | "source": [ 34 | "import os\n", 35 | "import openai\n", 36 | "from dotenv import load_dotenv, find_dotenv\n", 37 | "_ = load_dotenv(find_dotenv()) # read local .env file\n", 38 | "\n", 39 | "openai.api_key = os.getenv('OPENAI_API_KEY')" 40 | ] 41 | }, 42 | { 43 | "cell_type": "code", 44 | "execution_count": null, 45 | "metadata": {}, 46 | "outputs": [], 47 | "source": [ 48 | "def get_completion(prompt, model=\"gpt-3.5-turbo\"):\n", 49 | " messages = [{\"role\": \"user\", \"content\": prompt}]\n", 50 | " response = openai.ChatCompletion.create(\n", 51 | " model=model,\n", 52 | " messages=messages,\n", 53 | " temperature=0, # this is the degree of randomness of the model's output\n", 54 | " )\n", 55 | " return response.choices[0].message[\"content\"]\n", 56 | "\n", 57 | "def get_completion_from_messages(messages, model=\"gpt-3.5-turbo\", temperature=0):\n", 58 | " response = openai.ChatCompletion.create(\n", 59 | " model=model,\n", 60 | " messages=messages,\n", 61 | " temperature=temperature, # this is the degree of randomness of the model's output\n", 62 | " )\n", 63 | "# print(str(response.choices[0].message))\n", 64 | " return response.choices[0].message[\"content\"]" 65 | ] 66 | }, 67 | { 68 | "cell_type": "code", 69 | "execution_count": null, 70 | "metadata": {}, 71 | "outputs": [], 72 | "source": [ 73 | "messages = [ \n", 74 | "{'role':'system', 'content':'You are an assistant that speaks like Shakespeare.'}, \n", 75 | "{'role':'user', 'content':'tell me a joke'}, \n", 76 | "{'role':'assistant', 'content':'Why did the chicken cross the road'}, \n", 77 | "{'role':'user', 'content':'I don\\'t know'} ]" 78 | ] 79 | }, 80 | { 81 | "cell_type": "code", 82 | "execution_count": null, 83 | "metadata": {}, 84 | "outputs": [], 85 | "source": [ 86 | "response = get_completion_from_messages(messages, temperature=1)\n", 87 | "print(response)" 88 | ] 89 | }, 90 | { 91 | "cell_type": "code", 92 | "execution_count": null, 93 | "metadata": {}, 94 | "outputs": [], 95 | "source": [ 96 | "messages = [ \n", 97 | "{'role':'system', 'content':'You are friendly chatbot.'}, \n", 98 | "{'role':'user', 'content':'Hi, my name is Isa'} ]\n", 99 | "response = get_completion_from_messages(messages, temperature=1)\n", 100 | "print(response)" 101 | ] 102 | }, 103 | { 104 | "cell_type": "code", 105 | "execution_count": null, 106 | "metadata": {}, 107 | "outputs": [], 108 | "source": [ 109 | "messages = [ \n", 110 | "{'role':'system', 'content':'You are friendly chatbot.'}, \n", 111 | "{'role':'user', 'content':'Yes, can you remind me, What is my name?'} ]\n", 112 | "response = get_completion_from_messages(messages, temperature=1)\n", 113 | "print(response)" 114 | ] 115 | }, 116 | { 117 | "cell_type": "code", 118 | "execution_count": null, 119 | "metadata": {}, 120 | "outputs": [], 121 | "source": [ 122 | "messages = [ \n", 123 | "{'role':'system', 'content':'You are friendly chatbot.'},\n", 124 | "{'role':'user', 'content':'Hi, my name is Isa'},\n", 125 | "{'role':'assistant', 'content': \"Hi Isa! It's nice to meet you. \\\n", 126 | "Is there anything I can help you with today?\"},\n", 127 | "{'role':'user', 'content':'Yes, you can remind me, What is my name?'} ]\n", 128 | "response = get_completion_from_messages(messages, temperature=1)\n", 129 | "print(response)" 130 | ] 131 | }, 132 | { 133 | "attachments": {}, 134 | "cell_type": "markdown", 135 | "metadata": {}, 136 | "source": [ 137 | "# OrderBot\n", 138 | "We can automate the collection of user prompts and assistant responses to build a OrderBot. The OrderBot will take orders at a pizza restaurant. " 139 | ] 140 | }, 141 | { 142 | "cell_type": "code", 143 | "execution_count": null, 144 | "metadata": {}, 145 | "outputs": [], 146 | "source": [ 147 | "!pip install panel jupyter_bokeh" 148 | ] 149 | }, 150 | { 151 | "cell_type": "code", 152 | "execution_count": null, 153 | "metadata": {}, 154 | "outputs": [], 155 | "source": [ 156 | "def collect_messages(_):\n", 157 | " prompt = inp.value_input\n", 158 | " inp.value = ''\n", 159 | " context.append({'role':'user', 'content':f\"{prompt}\"})\n", 160 | " response = get_completion_from_messages(context) \n", 161 | " context.append({'role':'assistant', 'content':f\"{response}\"})\n", 162 | " panels.append(\n", 163 | " pn.Row('User:', pn.pane.Markdown(prompt, width=600)))\n", 164 | " panels.append(\n", 165 | " pn.Row('Assistant:', pn.pane.Markdown(response, width=600, style={'background-color': '#F6F6F6'})))\n", 166 | " \n", 167 | " return pn.Column(*panels)" 168 | ] 169 | }, 170 | { 171 | "cell_type": "code", 172 | "execution_count": null, 173 | "metadata": {}, 174 | "outputs": [], 175 | "source": [ 176 | "import panel as pn # GUI\n", 177 | "pn.extension()\n", 178 | "\n", 179 | "panels = [] # collect display \n", 180 | "\n", 181 | "context = [ {'role':'system', 'content':\"\"\"\n", 182 | "You are OrderBot, an automated service to collect orders for a pizza restaurant. \\\n", 183 | "You first greet the customer, then collects the order, \\\n", 184 | "and then asks if it's a pickup or delivery. \\\n", 185 | "You wait to collect the entire order, then summarize it and check for a final \\\n", 186 | "time if the customer wants to add anything else. \\\n", 187 | "If it's a delivery, you ask for an address. \\\n", 188 | "Finally you collect the payment.\\\n", 189 | "Make sure to clarify all options, extras and sizes to uniquely \\\n", 190 | "identify the item from the menu.\\\n", 191 | "You respond in a short, very conversational friendly style. \\\n", 192 | "The menu includes \\\n", 193 | "pepperoni pizza 12.95, 10.00, 7.00 \\\n", 194 | "cheese pizza 10.95, 9.25, 6.50 \\\n", 195 | "eggplant pizza 11.95, 9.75, 6.75 \\\n", 196 | "fries 4.50, 3.50 \\\n", 197 | "greek salad 7.25 \\\n", 198 | "Toppings: \\\n", 199 | "extra cheese 2.00, \\\n", 200 | "mushrooms 1.50 \\\n", 201 | "sausage 3.00 \\\n", 202 | "canadian bacon 3.50 \\\n", 203 | "AI sauce 1.50 \\\n", 204 | "peppers 1.00 \\\n", 205 | "Drinks: \\\n", 206 | "coke 3.00, 2.00, 1.00 \\\n", 207 | "sprite 3.00, 2.00, 1.00 \\\n", 208 | "bottled water 5.00 \\\n", 209 | "\"\"\"} ] # accumulate messages\n", 210 | "\n", 211 | "\n", 212 | "inp = pn.widgets.TextInput(value=\"Hi\", placeholder='Enter text here…')\n", 213 | "button_conversation = pn.widgets.Button(name=\"Chat!\")\n", 214 | "\n", 215 | "interactive_conversation = pn.bind(collect_messages, button_conversation)\n", 216 | "\n", 217 | "dashboard = pn.Column(\n", 218 | " inp,\n", 219 | " pn.Row(button_conversation),\n", 220 | " pn.panel(interactive_conversation, loading_indicator=True, height=300),\n", 221 | ")\n", 222 | "\n", 223 | "dashboard" 224 | ] 225 | }, 226 | { 227 | "cell_type": "code", 228 | "execution_count": null, 229 | "metadata": {}, 230 | "outputs": [], 231 | "source": [ 232 | "messages = context.copy()\n", 233 | "messages.append(\n", 234 | "{'role':'system', 'content':'create a json summary of the previous food order. Itemize the price for each item\\\n", 235 | " The fields should be 1) pizza, include size 2) list of toppings 3) list of drinks, include size 4) list of sides include size 5)total price '}, \n", 236 | ")\n", 237 | " #The fields should be 1) pizza, price 2) list of toppings 3) list of drinks, include size include price 4) list of sides include size include price, 5)total price '}, \n", 238 | "\n", 239 | "response = get_completion_from_messages(messages, temperature=0)\n", 240 | "print(response)" 241 | ] 242 | }, 243 | { 244 | "cell_type": "code", 245 | "execution_count": null, 246 | "metadata": {}, 247 | "outputs": [], 248 | "source": [] 249 | }, 250 | { 251 | "cell_type": "markdown", 252 | "metadata": {}, 253 | "source": [] 254 | } 255 | ], 256 | "metadata": { 257 | "kernelspec": { 258 | "display_name": "gpt_index", 259 | "language": "python", 260 | "name": "python3" 261 | }, 262 | "language_info": { 263 | "codemirror_mode": { 264 | "name": "ipython", 265 | "version": 3 266 | }, 267 | "file_extension": ".py", 268 | "mimetype": "text/x-python", 269 | "name": "python", 270 | "nbconvert_exporter": "python", 271 | "pygments_lexer": "ipython3", 272 | "version": "3.8.16" 273 | }, 274 | "orig_nbformat": 4 275 | }, 276 | "nbformat": 4, 277 | "nbformat_minor": 2 278 | } 279 | -------------------------------------------------------------------------------- /english_subtitles/prompt_eng_04_summarizing_english.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:09,680 3 | There's so much text in today's world, pretty much none of us have enough time to read all 4 | 5 | 2 6 | 00:00:09,680 --> 00:00:12,760 7 | the things we wish we had time to 8 | 9 | 3 10 | 00:00:12,760 --> 00:00:16,920 11 | So one of the most exciting applications I've seen of large language models is to use it 12 | 13 | 4 14 | 00:00:16,920 --> 00:00:18,660 15 | to summarize text 16 | 17 | 5 18 | 00:00:18,660 --> 00:00:24,480 19 | And this is something that I'm seeing multiple teams build into multiple software applications 20 | 21 | 6 22 | 00:00:24,480 --> 00:00:26,880 23 | You can do this in the ChatGPT web interface 24 | 25 | 7 26 | 00:00:26,880 --> 00:00:31,000 27 | I do this all the time to summarize articles so I can just kind of read the content of 28 | 29 | 8 30 | 00:00:31,000 --> 00:00:33,000 31 | many more articles than I previously could 32 | 33 | 9 34 | 00:00:33,000 --> 00:00:37,880 35 | And if you want to do this more programmatically, you'll see how to in this lesson 36 | 37 | 10 38 | 00:00:37,880 --> 00:00:43,520 39 | So that's let's dig into the code to see how you could use this yourself to summarize text 40 | 41 | 11 42 | 00:00:43,520 --> 00:00:48,280 43 | So let's start off with the same starter code as you saw before of import openai, load the 44 | 45 | 12 46 | 00:00:48,280 --> 00:00:52,620 47 | API key and here's that get_completion helper function 48 | 49 | 13 50 | 00:00:52,620 --> 00:01:00,640 51 | I'm going to use as the running example, the task of summarizing this product review 52 | 53 | 14 54 | 00:01:00,640 --> 00:01:04,640 55 | Got this panda plush toy for my daughter's birthday 56 | 57 | 15 58 | 00:01:04,640 --> 00:01:06,320 59 | and so on and so on 60 | 61 | 16 62 | 00:01:06,320 --> 00:01:13,640 63 | If you're building an e commerce website, and there's just a large volume of reviews 64 | 65 | 17 66 | 00:01:13,640 --> 00:01:18,680 67 | having a tool to summarize the lengthy reviews could give you a way to very quickly glance 68 | 69 | 18 70 | 00:01:18,680 --> 00:01:24,200 71 | over more reviews to get a better sense of what all your customers are thinking 72 | 73 | 19 74 | 00:01:24,200 --> 00:01:27,280 75 | So here's a prompt for generating a summary 76 | 77 | 20 78 | 00:01:27,280 --> 00:01:30,600 79 | Your task is generate a short summary of a product review from e commerce website, summarize 80 | 81 | 21 82 | 00:01:30,600 --> 00:01:36,800 83 | review below, and so on, in at most 30 words 84 | 85 | 22 86 | 00:01:36,800 --> 00:01:42,400 87 | And so this is soft and cute panda plush toy loved by daughter, but small for the price 88 | 89 | 23 90 | 00:01:42,400 --> 00:01:43,400 91 | arrived early 92 | 93 | 24 94 | 00:01:43,400 --> 00:01:44,400 95 | Not bad 96 | 97 | 25 98 | 00:01:44,400 --> 00:01:45,400 99 | It's a pretty good summary 100 | 101 | 26 102 | 00:01:45,400 --> 00:01:49,760 103 | And as you saw in the previous video, you can also play with things like controlling 104 | 105 | 27 106 | 00:01:49,760 --> 00:01:55,720 107 | the character count or the number of sentences to affect the length of this summary 108 | 109 | 28 110 | 00:01:55,720 --> 00:01:59,920 111 | Now sometimes when creating a summary, if you have a very specific purpose in mind for 112 | 113 | 29 114 | 00:01:59,920 --> 00:02:05,600 115 | the summary, for example, if you want to give feedback to the shipping department, you can 116 | 117 | 30 118 | 00:02:05,600 --> 00:02:11,200 119 | also modify the prompt to reflect that so that they can generate a summary that is more 120 | 121 | 31 122 | 00:02:11,200 --> 00:02:15,600 123 | applicable to one particular group in your business 124 | 125 | 32 126 | 00:02:15,600 --> 00:02:26,880 127 | So for example, if I add to give feedback to the shipping department, let's say I change 128 | 129 | 33 130 | 00:02:26,880 --> 00:02:35,800 131 | this to start to focus on any aspects that mention shipping and delivery of the product 132 | 133 | 34 134 | 00:02:35,800 --> 00:02:42,560 135 | And if I run this, then again, you get a summary, but instead of starting off with soft and 136 | 137 | 35 138 | 00:02:42,560 --> 00:02:49,120 139 | cute panda plush toy, it now focuses on the fact that it arrived a day earlier than expected 140 | 141 | 36 142 | 00:02:49,120 --> 00:02:53,600 143 | And then it still has, you know, other details 144 | 145 | 37 146 | 00:02:53,600 --> 00:02:59,640 147 | Or as another example, if we aren't trying to give feedback to the shipping department 148 | 149 | 38 150 | 00:02:59,640 --> 00:03:09,520 151 | but let's say we want to give feedback to the pricing department 152 | 153 | 39 154 | 00:03:09,520 --> 00:03:15,760 155 | So the pricing department is responsible for determining the price of the product 156 | 157 | 40 158 | 00:03:15,760 --> 00:03:22,960 159 | And I'm going to tell it to focus on any aspects that are relevant to the price and perceived 160 | 161 | 41 162 | 00:03:22,960 --> 00:03:25,240 163 | value 164 | 165 | 42 166 | 00:03:25,240 --> 00:03:31,520 167 | Then this generates a different summary that it says 168 | 169 | 43 170 | 00:03:31,520 --> 00:03:33,480 171 | maybe the price may be too high for a size 172 | 173 | 44 174 | 00:03:33,480 --> 00:03:38,640 175 | Now in the summaries that I've generated for the shipping department or the pricing department 176 | 177 | 45 178 | 00:03:38,640 --> 00:03:44,400 179 | it focuses a bit more on information relevant to those specific departments 180 | 181 | 46 182 | 00:03:44,400 --> 00:03:50,000 183 | And in fact, feel free to pause the video now and maybe ask it to generate information 184 | 185 | 47 186 | 00:03:50,000 --> 00:03:55,360 187 | for the product department responsible for the customer experience of the product or 188 | 189 | 48 190 | 00:03:55,360 --> 00:04:01,920 191 | for something else that you think might be interesting to an e-commerce site 192 | 193 | 49 194 | 00:04:01,920 --> 00:04:07,000 195 | But in these summaries, even though it generated the information relevant to shipping, it had 196 | 197 | 50 198 | 00:04:07,000 --> 00:04:12,240 199 | some other information too, which you could decide may or may not be helpful 200 | 201 | 51 202 | 00:04:12,240 --> 00:04:20,120 203 | So depending on how you want to summarize it, you can also ask it to extract information 204 | 205 | 52 206 | 00:04:20,120 --> 00:04:22,960 207 | rather than summarize it 208 | 209 | 53 210 | 00:04:22,960 --> 00:04:27,920 211 | So here's a prompt that says you're tasked to extract relevant information to give feedback 212 | 213 | 54 214 | 00:04:27,920 --> 00:04:29,200 215 | to the shipping department 216 | 217 | 55 218 | 00:04:29,200 --> 00:04:35,440 219 | And now it just says, product arrived a day earlier than expected without all of the other 220 | 221 | 56 222 | 00:04:35,440 --> 00:04:41,240 223 | information, which was also helpful in a general summary, but less specific to the shipping 224 | 225 | 57 226 | 00:04:41,240 --> 00:04:45,000 227 | department if all it wants to know is what happened with the shipping 228 | 229 | 58 230 | 00:04:45,000 --> 00:04:50,600 231 | Lastly, let me just share with you a concrete example for how to use this in a workflow 232 | 233 | 59 234 | 00:04:50,600 --> 00:04:56,760 235 | to help summarize multiple reviews to make them easier to read 236 | 237 | 60 238 | 00:04:56,760 --> 00:04:59,920 239 | So here are a few reviews 240 | 241 | 61 242 | 00:04:59,920 --> 00:05:03,920 243 | This is kind of long, but here's a second review for a standing lamp 244 | 245 | 62 246 | 00:05:03,920 --> 00:05:04,920 247 | need a lamp on a bedroom 248 | 249 | 63 250 | 00:05:04,920 --> 00:05:07,280 251 | Here's a third review for an electric toothbrush 252 | 253 | 64 254 | 00:05:07,280 --> 00:05:09,600 255 | My dental hygienist recommended 256 | 257 | 65 258 | 00:05:09,600 --> 00:05:12,640 259 | Kind of a long review about the electric toothbrush 260 | 261 | 66 262 | 00:05:12,640 --> 00:05:17,920 263 | This is a review for a blender when it said, so said 17 piece system on seasonal sale and 264 | 265 | 67 266 | 00:05:17,920 --> 00:05:18,920 267 | so on and so on 268 | 269 | 68 270 | 00:05:18,920 --> 00:05:20,280 271 | This is actually a lot of text 272 | 273 | 69 274 | 00:05:20,280 --> 00:05:23,280 275 | If you want, feel free to pause the video and read through all this text 276 | 277 | 70 278 | 00:05:23,280 --> 00:05:28,480 279 | But what if you want to know what these reviewers wrote without having to stop and read all 280 | 281 | 71 282 | 00:05:28,480 --> 00:05:29,480 283 | this in detail? 284 | 285 | 72 286 | 00:05:29,480 --> 00:05:41,600 287 | So I'm going to set review_1 to be just the product review that we had up there 288 | 289 | 73 290 | 00:05:41,600 --> 00:05:45,320 291 | I'm going to put all of these reviews into a list 292 | 293 | 74 294 | 00:05:45,320 --> 00:05:53,800 295 | And now if I implement a for loop over the reviews 296 | 297 | 75 298 | 00:05:53,800 --> 00:06:01,240 299 | So here's my prompt and here I've asked it to summarize it in at most 20 words 300 | 301 | 76 302 | 00:06:01,240 --> 00:06:07,320 303 | Then let's have it get the response and print it out 304 | 305 | 77 306 | 00:06:07,320 --> 00:06:10,760 307 | And let's run that 308 | 309 | 78 310 | 00:06:10,760 --> 00:06:16,360 311 | And it prints out the first review is that PandaToy review, summary review of the lamp 312 | 313 | 79 314 | 00:06:16,360 --> 00:06:21,880 315 | summary review of the toothbrush, and then the blender 316 | 317 | 80 318 | 00:06:21,880 --> 00:06:29,800 319 | And so if you have a website where you have hundreds of reviews, you can imagine how you 320 | 321 | 81 322 | 00:06:29,800 --> 00:06:35,360 323 | might use this to build a dashboard, to take huge numbers of reviews, generate short summaries 324 | 325 | 82 326 | 00:06:35,360 --> 00:06:40,760 327 | of them so that you or someone else can browse the reviews much more quickly 328 | 329 | 83 330 | 00:06:40,760 --> 00:06:46,200 331 | And then if they wish, maybe click in to see the original longer review 332 | 333 | 84 334 | 00:06:46,200 --> 00:06:51,640 335 | And this can help you efficiently get a better sense of what all of your customers are thinking 336 | 337 | 85 338 | 00:06:51,640 --> 00:06:55,640 339 | All right, so that's it for summarizing 340 | 341 | 86 342 | 00:06:55,640 --> 00:07:00,960 343 | And I hope that you can picture if you have any applications with many pieces of text 344 | 345 | 87 346 | 00:07:00,960 --> 00:07:05,680 347 | how you can use prompts like these to summarize them to help people quickly get a sense of 348 | 349 | 88 350 | 00:07:05,680 --> 00:07:10,240 351 | what's in the text, the many pieces of text, and perhaps optionally dig in more 352 | 353 | 89 354 | 00:07:10,240 --> 00:07:12,520 355 | if they wish 356 | 357 | 90 358 | 00:07:12,520 --> 00:07:17,640 359 | In the next video, we'll look at another capability of large language models, which is to make 360 | 361 | 91 362 | 00:07:17,640 --> 00:07:19,880 363 | inferences using text 364 | 365 | 92 366 | 00:07:19,880 --> 00:07:24,280 367 | For example, what if you had, again, product reviews, and you wanted to very quickly get 368 | 369 | 93 370 | 00:07:24,280 --> 00:07:29,680 371 | a sense of which product reviews have a positive or a negative sentiment? 372 | 373 | 94 374 | 00:07:29,680 --> 00:07:49,840 375 | Let's take a look at how to do that in the next video 376 | 377 | -------------------------------------------------------------------------------- /course_notebooks/l3-iterative-prompt-development.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "attachments": {}, 5 | "cell_type": "markdown", 6 | "metadata": {}, 7 | "source": [ 8 | "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/GitHubDaily/ChatGPT-Prompt-Engineering-for-Developers-in-Chinese/blob/master/assets/3-Iterative/l3-iterative-prompt-development.ipynb)\n", 9 | "# Iterative Prompt Develelopment\n", 10 | "In this lesson, you'll iteratively analyze and refine your prompts to generate marketing copy from a product fact sheet.\n", 11 | "\n", 12 | "## Setup" 13 | ] 14 | }, 15 | { 16 | "cell_type": "code", 17 | "execution_count": null, 18 | "metadata": {}, 19 | "outputs": [], 20 | "source": [ 21 | "# Basic congfig \n", 22 | "# Install basic package and set key\n", 23 | "!pip install openai\n", 24 | "!export OPENAI_API_KEY='sk-...'" 25 | ] 26 | }, 27 | { 28 | "cell_type": "code", 29 | "execution_count": null, 30 | "metadata": {}, 31 | "outputs": [], 32 | "source": [ 33 | "import openai\n", 34 | "import os\n", 35 | "\n", 36 | "from dotenv import load_dotenv, find_dotenv\n", 37 | "_ = load_dotenv(find_dotenv()) # read local .env file\n", 38 | "\n", 39 | "openai.api_key = os.getenv('OPENAI_API_KEY')" 40 | ] 41 | }, 42 | { 43 | "cell_type": "code", 44 | "execution_count": null, 45 | "metadata": {}, 46 | "outputs": [], 47 | "source": [ 48 | "def get_completion(prompt, model=\"gpt-3.5-turbo\"):\n", 49 | " messages = [{\"role\": \"user\", \"content\": prompt}]\n", 50 | " response = openai.ChatCompletion.create(\n", 51 | " model=model,\n", 52 | " messages=messages,\n", 53 | " temperature=0, # this is the degree of randomness of the model's output\n", 54 | " )\n", 55 | " return response.choices[0].message[\"content\"]" 56 | ] 57 | }, 58 | { 59 | "attachments": {}, 60 | "cell_type": "markdown", 61 | "metadata": {}, 62 | "source": [ 63 | "Generate a marketing product description from a product fact sheet" 64 | ] 65 | }, 66 | { 67 | "cell_type": "code", 68 | "execution_count": null, 69 | "metadata": {}, 70 | "outputs": [], 71 | "source": [ 72 | "fact_sheet_chair = \"\"\"\n", 73 | "OVERVIEW\n", 74 | "- Part of a beautiful family of mid-century inspired office furniture, \n", 75 | "including filing cabinets, desks, bookcases, meeting tables, and more.\n", 76 | "- Several options of shell color and base finishes.\n", 77 | "- Available with plastic back and front upholstery (SWC-100) \n", 78 | "or full upholstery (SWC-110) in 10 fabric and 6 leather options.\n", 79 | "- Base finish options are: stainless steel, matte black, \n", 80 | "gloss white, or chrome.\n", 81 | "- Chair is available with or without armrests.\n", 82 | "- Suitable for home or business settings.\n", 83 | "- Qualified for contract use.\n", 84 | "\n", 85 | "CONSTRUCTION\n", 86 | "- 5-wheel plastic coated aluminum base.\n", 87 | "- Pneumatic chair adjust for easy raise/lower action.\n", 88 | "\n", 89 | "DIMENSIONS\n", 90 | "- WIDTH 53 CM | 20.87”\n", 91 | "- DEPTH 51 CM | 20.08”\n", 92 | "- HEIGHT 80 CM | 31.50”\n", 93 | "- SEAT HEIGHT 44 CM | 17.32”\n", 94 | "- SEAT DEPTH 41 CM | 16.14”\n", 95 | "\n", 96 | "OPTIONS\n", 97 | "- Soft or hard-floor caster options.\n", 98 | "- Two choices of seat foam densities: \n", 99 | " medium (1.8 lb/ft3) or high (2.8 lb/ft3)\n", 100 | "- Armless or 8 position PU armrests \n", 101 | "\n", 102 | "MATERIALS\n", 103 | "SHELL BASE GLIDER\n", 104 | "- Cast Aluminum with modified nylon PA6/PA66 coating.\n", 105 | "- Shell thickness: 10 mm.\n", 106 | "SEAT\n", 107 | "- HD36 foam\n", 108 | "\n", 109 | "COUNTRY OF ORIGIN\n", 110 | "- Italy\n", 111 | "\"\"\"" 112 | ] 113 | }, 114 | { 115 | "cell_type": "code", 116 | "execution_count": null, 117 | "metadata": {}, 118 | "outputs": [], 119 | "source": [ 120 | "prompt = f\"\"\"\n", 121 | "Your task is to help a marketing team create a \n", 122 | "description for a retail website of a product based \n", 123 | "on a technical fact sheet.\n", 124 | "\n", 125 | "Write a product description based on the information \n", 126 | "provided in the technical specifications delimited by \n", 127 | "triple backticks.\n", 128 | "\n", 129 | "Technical specifications: ```{fact_sheet_chair}```\n", 130 | "\"\"\"\n", 131 | "response = get_completion(prompt)\n", 132 | "print(response)\n" 133 | ] 134 | }, 135 | { 136 | "attachments": {}, 137 | "cell_type": "markdown", 138 | "metadata": {}, 139 | "source": [ 140 | "## Issue 1: The text is too long \n", 141 | "- Limit the number of words/sentences/characters." 142 | ] 143 | }, 144 | { 145 | "cell_type": "code", 146 | "execution_count": null, 147 | "metadata": {}, 148 | "outputs": [], 149 | "source": [ 150 | "prompt = f\"\"\"\n", 151 | "Your task is to help a marketing team create a \n", 152 | "description for a retail website of a product based \n", 153 | "on a technical fact sheet.\n", 154 | "\n", 155 | "Write a product description based on the information \n", 156 | "provided in the technical specifications delimited by \n", 157 | "triple backticks.\n", 158 | "\n", 159 | "Use at most 50 words.\n", 160 | "\n", 161 | "Technical specifications: ```{fact_sheet_chair}```\n", 162 | "\"\"\"\n", 163 | "response = get_completion(prompt)\n", 164 | "print(response)" 165 | ] 166 | }, 167 | { 168 | "cell_type": "code", 169 | "execution_count": null, 170 | "metadata": {}, 171 | "outputs": [], 172 | "source": [ 173 | "len(response)" 174 | ] 175 | }, 176 | { 177 | "attachments": {}, 178 | "cell_type": "markdown", 179 | "metadata": {}, 180 | "source": [ 181 | "## Issue 2. Text focuses on the wrong details\n", 182 | "- Ask it to focus on the aspects that are relevant to the intended audience." 183 | ] 184 | }, 185 | { 186 | "cell_type": "code", 187 | "execution_count": null, 188 | "metadata": {}, 189 | "outputs": [], 190 | "source": [ 191 | "prompt = f\"\"\"\n", 192 | "Your task is to help a marketing team create a \n", 193 | "description for a retail website of a product based \n", 194 | "on a technical fact sheet.\n", 195 | "\n", 196 | "Write a product description based on the information \n", 197 | "provided in the technical specifications delimited by \n", 198 | "triple backticks.\n", 199 | "\n", 200 | "The description is intended for furniture retailers, \n", 201 | "so should be technical in nature and focus on the \n", 202 | "materials the product is constructed from.\n", 203 | "\n", 204 | "Use at most 50 words.\n", 205 | "\n", 206 | "Technical specifications: ```{fact_sheet_chair}```\n", 207 | "\"\"\"\n", 208 | "response = get_completion(prompt)\n", 209 | "print(response) " 210 | ] 211 | }, 212 | { 213 | "cell_type": "code", 214 | "execution_count": null, 215 | "metadata": {}, 216 | "outputs": [], 217 | "source": [ 218 | "prompt = f\"\"\"\n", 219 | "Your task is to help a marketing team create a \n", 220 | "description for a retail website of a product based \n", 221 | "on a technical fact sheet.\n", 222 | "\n", 223 | "Write a product description based on the information \n", 224 | "provided in the technical specifications delimited by \n", 225 | "triple backticks.\n", 226 | "\n", 227 | "The description is intended for furniture retailers, \n", 228 | "so should be technical in nature and focus on the \n", 229 | "materials the product is constructed from.\n", 230 | "\n", 231 | "At the end of the description, include every 7-character \n", 232 | "Product ID in the technical specification.\n", 233 | "\n", 234 | "Use at most 50 words.\n", 235 | "\n", 236 | "Technical specifications: ```{fact_sheet_chair}```\n", 237 | "\"\"\"\n", 238 | "response = get_completion(prompt)\n", 239 | "print(response) " 240 | ] 241 | }, 242 | { 243 | "attachments": {}, 244 | "cell_type": "markdown", 245 | "metadata": {}, 246 | "source": [ 247 | "## Issue 3. Description needs a table of dimensions\n", 248 | "- Ask it to extract information and organize it in a table." 249 | ] 250 | }, 251 | { 252 | "cell_type": "code", 253 | "execution_count": null, 254 | "metadata": {}, 255 | "outputs": [], 256 | "source": [ 257 | "prompt = f\"\"\"\n", 258 | "Your task is to help a marketing team create a \n", 259 | "description for a retail website of a product based \n", 260 | "on a technical fact sheet.\n", 261 | "\n", 262 | "Write a product description based on the information \n", 263 | "provided in the technical specifications delimited by \n", 264 | "triple backticks.\n", 265 | "\n", 266 | "The description is intended for furniture retailers, \n", 267 | "so should be technical in nature and focus on the \n", 268 | "materials the product is constructed from.\n", 269 | "\n", 270 | "At the end of the description, include every 7-character \n", 271 | "Product ID in the technical specification.\n", 272 | "\n", 273 | "After the description, include a table that gives the \n", 274 | "product's dimensions. The table should have two columns.\n", 275 | "In the first column include the name of the dimension. \n", 276 | "In the second column include the measurements in inches only.\n", 277 | "\n", 278 | "Give the table the title 'Product Dimensions'.\n", 279 | "\n", 280 | "Format everything as HTML that can be used in a website. \n", 281 | "Place the description in a
element.\n", 282 | "\n", 283 | "Technical specifications: ```{fact_sheet_chair}```\n", 284 | "\"\"\"\n", 285 | "\n", 286 | "response = get_completion(prompt)\n", 287 | "print(response)" 288 | ] 289 | }, 290 | { 291 | "attachments": {}, 292 | "cell_type": "markdown", 293 | "metadata": {}, 294 | "source": [ 295 | "## Load Python libraries to view HTML" 296 | ] 297 | }, 298 | { 299 | "cell_type": "code", 300 | "execution_count": null, 301 | "metadata": {}, 302 | "outputs": [], 303 | "source": [ 304 | "from IPython.display import display, HTML" 305 | ] 306 | }, 307 | { 308 | "cell_type": "code", 309 | "execution_count": null, 310 | "metadata": {}, 311 | "outputs": [], 312 | "source": [ 313 | "display(HTML(response))" 314 | ] 315 | } 316 | ], 317 | "metadata": { 318 | "kernelspec": { 319 | "display_name": "gpt_index", 320 | "language": "python", 321 | "name": "python3" 322 | }, 323 | "language_info": { 324 | "codemirror_mode": { 325 | "name": "ipython", 326 | "version": 3 327 | }, 328 | "file_extension": ".py", 329 | "mimetype": "text/x-python", 330 | "name": "python", 331 | "nbconvert_exporter": "python", 332 | "pygments_lexer": "ipython3", 333 | "version": "3.8.16" 334 | }, 335 | "orig_nbformat": 4 336 | }, 337 | "nbformat": 4, 338 | "nbformat_minor": 2 339 | } 340 | -------------------------------------------------------------------------------- /english_subtitles/prompt_eng_01_intro_english.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:09,000 3 | Welcome to this course on ChatGPT Prompt Engineering for Developers 4 | 5 | 2 6 | 00:00:09,000 --> 00:00:14,000 7 | I'm thrilled to have with me Iza Fulford to teach this along with me 8 | 9 | 3 10 | 00:00:14,000 --> 00:00:18,000 11 | She is a member of the technical staff of OpenAI and had built the popular 12 | 13 | 4 14 | 00:00:18,000 --> 00:00:23,000 15 | ChatGPT Retrieval plugin and a large part of her work has been teaching people 16 | 17 | 5 18 | 00:00:23,000 --> 00:00:27,000 19 | how to use LLM or large language model technology in products 20 | 21 | 6 22 | 00:00:27,000 --> 00:00:31,000 23 | She's also contributed to the OpenAI cookbook that teaches people prompting 24 | 25 | 7 26 | 00:00:31,000 --> 00:00:32,000 27 | So thrilled to have you with you 28 | 29 | 8 30 | 00:00:32,000 --> 00:00:37,000 31 | And I'm thrilled to be here and share some prompting best practices with you all 32 | 33 | 9 34 | 00:00:37,000 --> 00:00:42,000 35 | So there's been a lot of material on the internet for prompting with articles 36 | 37 | 10 38 | 00:00:42,000 --> 00:00:45,000 39 | like 30 prompts everyone has to know 40 | 41 | 11 42 | 00:00:45,000 --> 00:00:50,000 43 | A lot of that has been focused on the ChatGPT web user interface 44 | 45 | 12 46 | 00:00:50,000 --> 00:00:54,000 47 | which many people are using to do specific and often one-off tasks 48 | 49 | 13 50 | 00:00:54,000 --> 00:01:00,000 51 | But I think the power of LLMs, large language models as a developer too 52 | 53 | 14 54 | 00:01:00,000 --> 00:01:04,000 55 | that is using API calls to LLMs to quickly build software applications 56 | 57 | 15 58 | 00:01:04,000 --> 00:01:08,000 59 | I think that is still very underappreciated 60 | 61 | 16 62 | 00:01:08,000 --> 00:01:12,000 63 | In fact, my team at AI Fund, which is a sister company to Deep Learning.AI 64 | 65 | 17 66 | 00:01:12,000 --> 00:01:16,000 67 | has been working with many startups on applying these technologies 68 | 69 | 18 70 | 00:01:16,000 --> 00:01:18,000 71 | to many different applications 72 | 73 | 19 74 | 00:01:18,000 --> 00:01:23,000 75 | And it's been exciting to see what LLM APIs can enable developers 76 | 77 | 20 78 | 00:01:23,000 --> 00:01:25,000 79 | to very quickly build 80 | 81 | 21 82 | 00:01:25,000 --> 00:01:29,000 83 | So in this course, we'll share with you some of the possibilities 84 | 85 | 22 86 | 00:01:29,000 --> 00:01:34,000 87 | for what you can do as well as best practices for how you can do them 88 | 89 | 23 90 | 00:01:34,000 --> 00:01:36,000 91 | There's a lot of material to cover 92 | 93 | 24 94 | 00:01:36,000 --> 00:01:41,000 95 | First, you'll learn some prompting best practices for software development 96 | 97 | 25 98 | 00:01:41,000 --> 00:01:45,000 99 | Then we'll cover some common use cases, summarizing, inferring 100 | 101 | 26 102 | 00:01:45,000 --> 00:01:50,000 103 | transforming, expanding, and then you'll build a chatbot using an LLM 104 | 105 | 27 106 | 00:01:50,000 --> 00:01:53,000 107 | We hope that this will spark your imagination about new applications 108 | 109 | 28 110 | 00:01:53,000 --> 00:01:55,000 111 | that you can build 112 | 113 | 29 114 | 00:01:55,000 --> 00:01:58,000 115 | So in the development of large language models or LLMs 116 | 117 | 30 118 | 00:01:58,000 --> 00:02:02,000 119 | there have been broadly two types of LLMs, which I'm going to refer to 120 | 121 | 31 122 | 00:02:02,000 --> 00:02:06,000 123 | as base LLMs and instruction-tuned LLMs 124 | 125 | 32 126 | 00:02:06,000 --> 00:02:11,000 127 | So base LLM has been trained to predict the next word based on text training data 128 | 129 | 33 130 | 00:02:11,000 --> 00:02:15,000 131 | often trained on a large amount of data from the internet and other sources 132 | 133 | 34 134 | 00:02:15,000 --> 00:02:19,000 135 | to figure out what's the next most likely word to follow 136 | 137 | 35 138 | 00:02:19,000 --> 00:02:24,000 139 | So for example, if you were to prompt this, once upon a time there was a unicorn 140 | 141 | 36 142 | 00:02:24,000 --> 00:02:28,000 143 | it may complete this, that is, it may predict the next several words are 144 | 145 | 37 146 | 00:02:28,000 --> 00:02:31,000 147 | that live in a magical forest with all unicorn friends 148 | 149 | 38 150 | 00:02:31,000 --> 00:02:35,000 151 | But if you were to prompt this with what is the capital of France 152 | 153 | 39 154 | 00:02:35,000 --> 00:02:40,000 155 | then based on what articles on the internet might have 156 | 157 | 40 158 | 00:02:40,000 --> 00:02:44,000 159 | it's quite possible that the base LLM will complete this with 160 | 161 | 41 162 | 00:02:44,000 --> 00:02:48,000 163 | what is France's largest city, what is France's population, and so on 164 | 165 | 42 166 | 00:02:48,000 --> 00:02:52,000 167 | because articles on the internet could quite plausibly be lists of quiz questions 168 | 169 | 43 170 | 00:02:52,000 --> 00:02:55,000 171 | about the country of France 172 | 173 | 44 174 | 00:02:55,000 --> 00:03:00,000 175 | In contrast, an instruction-tuned LLM, which is where a lot of momentum 176 | 177 | 45 178 | 00:03:00,000 --> 00:03:04,000 179 | of LLM research and practice has been going 180 | 181 | 46 182 | 00:03:04,000 --> 00:03:08,000 183 | an instruction-tuned LLM has been trained to follow instructions 184 | 185 | 47 186 | 00:03:08,000 --> 00:03:11,000 187 | So if you were to ask it, what is the capital of France 188 | 189 | 48 190 | 00:03:11,000 --> 00:03:15,000 191 | it's much more likely to output something like the capital of France is Paris 192 | 193 | 49 194 | 00:03:15,000 --> 00:03:19,000 195 | So the way that instruction-tuned LLMs are typically trained is 196 | 197 | 50 198 | 00:03:19,000 --> 00:03:23,000 199 | you start off with a base LLM that's been trained on a huge amount of text data 200 | 201 | 51 202 | 00:03:23,000 --> 00:03:28,000 203 | and further train it, further fine-tune it with inputs and outputs 204 | 205 | 52 206 | 00:03:28,000 --> 00:03:32,000 207 | that are instructions and good attempts to follow those instructions 208 | 209 | 53 210 | 00:03:32,000 --> 00:03:36,000 211 | and then often further refine using a technique called RLHF 212 | 213 | 54 214 | 00:03:36,000 --> 00:03:41,000 215 | reinforcement learning from human feedback, to make the system better able 216 | 217 | 55 218 | 00:03:41,000 --> 00:03:43,000 219 | to be helpful and follow instructions 220 | 221 | 56 222 | 00:03:43,000 --> 00:03:47,000 223 | Because instruction-tuned LLMs have been trained to be helpful 224 | 225 | 57 226 | 00:03:47,000 --> 00:03:51,000 227 | honest, and harmless, so for example, they're less likely to output 228 | 229 | 58 230 | 00:03:51,000 --> 00:03:55,000 231 | problematic text, such as toxic outputs, compared to base LLM 232 | 233 | 59 234 | 00:03:55,000 --> 00:03:59,000 235 | a lot of the practical usage scenarios have been shifting 236 | 237 | 60 238 | 00:03:59,000 --> 00:04:01,000 239 | toward instruction-tuned LLMs 240 | 241 | 61 242 | 00:04:01,000 --> 00:04:04,000 243 | Some of the best practices you find on the internet may be more suited 244 | 245 | 62 246 | 00:04:04,000 --> 00:04:08,000 247 | for a base LLM, but for most practical applications today 248 | 249 | 63 250 | 00:04:08,000 --> 00:04:13,000 251 | we would recommend most people instead focus on instruction-tuned LLMs 252 | 253 | 64 254 | 00:04:13,000 --> 00:04:17,000 255 | which are easier to use, and also because of the work of OpenAI 256 | 257 | 65 258 | 00:04:17,000 --> 00:04:22,000 259 | and other LLM companies, becoming safer and more aligned 260 | 261 | 66 262 | 00:04:22,000 --> 00:04:27,000 263 | So this course will focus on best practices for instruction-tuned LLMs 264 | 265 | 67 266 | 00:04:27,000 --> 00:04:32,000 267 | which is what we recommend you use for most of your applications 268 | 269 | 68 270 | 00:04:32,000 --> 00:04:36,000 271 | Before moving on, I just want to acknowledge the team from OpenAI 272 | 273 | 69 274 | 00:04:36,000 --> 00:04:39,000 275 | and DeepLearning.ai that had contributed to the materials 276 | 277 | 70 278 | 00:04:39,000 --> 00:04:42,000 279 | that Isa and I will be presenting 280 | 281 | 71 282 | 00:04:42,000 --> 00:04:45,000 283 | I'm very grateful to Andrew Main, Joe Palermo, Boris Power 284 | 285 | 72 286 | 00:04:45,000 --> 00:04:49,000 287 | Ted Sanders, and Lilian Wang from OpenAI that were very involved 288 | 289 | 73 290 | 00:04:49,000 --> 00:04:53,000 291 | with us brainstorming materials, vetting the materials to put together 292 | 293 | 74 294 | 00:04:53,000 --> 00:04:55,000 295 | the curriculum for this short course 296 | 297 | 75 298 | 00:04:55,000 --> 00:04:58,000 299 | And I'm also grateful on the DeepLearning side for the work 300 | 301 | 76 302 | 00:04:58,000 --> 00:05:01,000 303 | of Jeff Ludwig, Eddie Hsu, and Tommy Nelson 304 | 305 | 77 306 | 00:05:01,000 --> 00:05:06,000 307 | So when you use an instruction-tuned LLM, think of giving instructions 308 | 309 | 78 310 | 00:05:06,000 --> 00:05:10,000 311 | to another person, say someone that's smart but doesn't know 312 | 313 | 79 314 | 00:05:10,000 --> 00:05:12,000 315 | the specifics of your task 316 | 317 | 80 318 | 00:05:12,000 --> 00:05:16,000 319 | So when an LLM doesn't work, sometimes it's because the instructions 320 | 321 | 81 322 | 00:05:16,000 --> 00:05:17,000 323 | weren't clear enough 324 | 325 | 82 326 | 00:05:17,000 --> 00:05:20,000 327 | For example, if you were to say, please write me something 328 | 329 | 83 330 | 00:05:20,000 --> 00:05:22,000 331 | about Alan Turing 332 | 333 | 84 334 | 00:05:22,000 --> 00:05:26,000 335 | Well, in addition to that, it can be helpful to be clear about 336 | 337 | 85 338 | 00:05:26,000 --> 00:05:30,000 339 | whether you want the text to focus on his scientific work 340 | 341 | 86 342 | 00:05:30,000 --> 00:05:34,000 343 | or his personal life or his role in history or something else 344 | 345 | 87 346 | 00:05:34,000 --> 00:05:39,000 347 | And if you specify what you want the tone of the text to be 348 | 349 | 88 350 | 00:05:39,000 --> 00:05:43,000 351 | should it take on the tone like a professional journalist would write 352 | 353 | 89 354 | 00:05:43,000 --> 00:05:46,000 355 | or is it more of a casual note that you dash off to a friend? 356 | 357 | 90 358 | 00:05:46,000 --> 00:05:47,000 359 | That holds 360 | 361 | 91 362 | 00:05:47,000 --> 00:05:49,000 363 | The LLM generates what you want 364 | 365 | 92 366 | 00:05:49,000 --> 00:05:52,000 367 | And of course, if you picture yourself asking, say 368 | 369 | 93 370 | 00:05:52,000 --> 00:05:56,000 371 | a fresh college graduate to carry out this task for you 372 | 373 | 94 374 | 00:05:56,000 --> 00:05:59,000 375 | if you can even specify what snippets of text they should read 376 | 377 | 95 378 | 00:05:59,000 --> 00:06:02,000 379 | in advance to write this text about Alan Turing 380 | 381 | 96 382 | 00:06:02,000 --> 00:06:06,000 383 | then that even better sets up that fresh college grad for success 384 | 385 | 97 386 | 00:06:06,000 --> 00:06:09,000 387 | to carry out this task for you 388 | 389 | 98 390 | 00:06:09,000 --> 00:06:13,000 391 | So in the next video, you see examples of how to be clear 392 | 393 | 99 394 | 00:06:13,000 --> 00:06:17,000 395 | and specific, which is an important principle of prompting LLMs 396 | 397 | 100 398 | 00:06:17,000 --> 00:06:21,000 399 | And you also learn from Isa a second principle of prompting 400 | 401 | 101 402 | 00:06:21,000 --> 00:06:24,000 403 | that is giving a LLM time to think 404 | 405 | 102 406 | 00:06:24,000 --> 00:06:29,000 407 | So with that, let's go on to the next video 408 | 409 | -------------------------------------------------------------------------------- /english_subtitles/prompt_eng_07_expanding_english.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:08,200 3 | Expanding is the task of taking a shorter piece of text 4 | 5 | 2 6 | 00:00:08,200 --> 00:00:11,140 7 | such as a set of instructions or a list of topics 8 | 9 | 3 10 | 00:00:11,140 --> 00:00:14,900 11 | and having the large language model generate a longer piece of text 12 | 13 | 4 14 | 00:00:14,900 --> 00:00:18,260 15 | such as an email or an essay about some topic 16 | 17 | 5 18 | 00:00:18,260 --> 00:00:20,100 19 | There are some great uses of this 20 | 21 | 6 22 | 00:00:20,100 --> 00:00:23,960 23 | such as if you use a large language model as a brainstorming partner 24 | 25 | 7 26 | 00:00:23,960 --> 00:00:26,300 27 | But I just also want to acknowledge that there's 28 | 29 | 8 30 | 00:00:26,300 --> 00:00:28,500 31 | some problematic use cases of this 32 | 33 | 9 34 | 00:00:28,500 --> 00:00:32,500 35 | such as if someone were to use it to generate a large amount of spam 36 | 37 | 10 38 | 00:00:32,500 --> 00:00:36,100 39 | So when you use these capabilities of a large language model 40 | 41 | 11 42 | 00:00:36,100 --> 00:00:41,340 43 | please use it only in a responsible way and in a way that helps people 44 | 45 | 12 46 | 00:00:41,340 --> 00:00:45,280 47 | In this video, we'll go through an example of how you can use 48 | 49 | 13 50 | 00:00:45,280 --> 00:00:49,900 51 | a language model to generate a personalized email based on some information 52 | 53 | 14 54 | 00:00:49,900 --> 00:00:53,260 55 | The email is kind of self-proclaimed to be from an AI bot 56 | 57 | 15 58 | 00:00:53,260 --> 00:00:56,140 59 | which as Andrew mentioned, is very important 60 | 61 | 16 62 | 00:00:56,140 --> 00:00:58,500 63 | We're also going to use another one of 64 | 65 | 17 66 | 00:00:58,500 --> 00:01:01,260 67 | the model's input parameters called temperature 68 | 69 | 18 70 | 00:01:01,260 --> 00:01:04,780 71 | and this allows you to vary the degree of 72 | 73 | 19 74 | 00:01:04,780 --> 00:01:08,760 75 | exploration and variety in the model's responses 76 | 77 | 20 78 | 00:01:08,760 --> 00:01:10,980 79 | So let's get into it 80 | 81 | 21 82 | 00:01:10,980 --> 00:01:12,740 83 | So before we get started 84 | 85 | 22 86 | 00:01:12,740 --> 00:01:15,780 87 | we're going to do the usual setup 88 | 89 | 23 90 | 00:01:15,780 --> 00:01:18,400 91 | So set up the OpenAI Python package 92 | 93 | 24 94 | 00:01:18,400 --> 00:01:22,060 95 | and then also define our helper function, get_completion 96 | 97 | 25 98 | 00:01:22,060 --> 00:01:26,460 99 | Now we're going to write a custom email response to a customer review 100 | 101 | 26 102 | 00:01:26,460 --> 00:01:29,940 103 | So given a customer review and the sentiment 104 | 105 | 27 106 | 00:01:29,940 --> 00:01:32,680 107 | we're going to generate a custom response 108 | 109 | 28 110 | 00:01:32,680 --> 00:01:37,100 111 | Now we're going to use the language model to generate 112 | 113 | 29 114 | 00:01:37,100 --> 00:01:41,900 115 | a custom email to a customer based on a customer review 116 | 117 | 30 118 | 00:01:41,900 --> 00:01:44,020 119 | and the sentiment of the review 120 | 121 | 31 122 | 00:01:44,020 --> 00:01:47,100 123 | So we've already extracted the sentiment 124 | 125 | 32 126 | 00:01:47,100 --> 00:01:52,740 127 | using the kind of prompts that we saw in the inferring video 128 | 129 | 33 130 | 00:01:52,740 --> 00:01:57,020 131 | and then this is the customer review for a blender 132 | 133 | 34 134 | 00:01:57,100 --> 00:02:02,740 135 | Now we're going to customize the reply based on the sentiment 136 | 137 | 35 138 | 00:02:02,740 --> 00:02:05,180 139 | So here the instruction is 140 | 141 | 36 142 | 00:02:05,180 --> 00:02:08,260 143 | you are a customer service AI assistant 144 | 145 | 37 146 | 00:02:08,260 --> 00:02:11,180 147 | Your task is to send an email reply to a valued customer 148 | 149 | 38 150 | 00:02:11,180 --> 00:02:14,540 151 | Given the customer email delimited by three backticks 152 | 153 | 39 154 | 00:02:14,540 --> 00:02:16,860 155 | generate a reply to thank the customer for their review 156 | 157 | 40 158 | 00:02:16,860 --> 00:02:18,780 159 | If the sentiment is positive or neutral 160 | 161 | 41 162 | 00:02:18,780 --> 00:02:20,460 163 | thank them for their review 164 | 165 | 42 166 | 00:02:20,460 --> 00:02:21,980 167 | If the sentiment is negative 168 | 169 | 43 170 | 00:02:21,980 --> 00:02:25,260 171 | apologize and suggest that they can reach out to customer service 172 | 173 | 44 174 | 00:02:25,260 --> 00:02:27,820 175 | Make sure to use specific details from the review 176 | 177 | 45 178 | 00:02:27,820 --> 00:02:29,820 179 | write in a concise and professional tone 180 | 181 | 46 182 | 00:02:29,820 --> 00:02:32,460 183 | and sign the email as AI customer agent 184 | 185 | 47 186 | 00:02:32,460 --> 00:02:34,820 187 | When you're using a language model 188 | 189 | 48 190 | 00:02:34,820 --> 00:02:37,500 191 | to generate text that you're going to show to a user 192 | 193 | 49 194 | 00:02:37,500 --> 00:02:40,860 195 | it's very important to have this kind of transparency 196 | 197 | 50 198 | 00:02:40,860 --> 00:02:45,700 199 | and let the user know that the text they're seeing was generated by AI 200 | 201 | 51 202 | 00:02:45,700 --> 00:02:49,780 203 | Then we'll just input the customer review and the review sentiment 204 | 205 | 52 206 | 00:02:49,780 --> 00:02:53,220 207 | Also note that this part isn't necessarily important 208 | 209 | 53 210 | 00:02:53,220 --> 00:02:57,180 211 | because we could actually use this prompt to also extract the review sentiment 212 | 213 | 54 214 | 00:02:57,180 --> 00:02:59,300 215 | and then in a follow-up step, write the email 216 | 217 | 55 218 | 00:02:59,300 --> 00:03:01,140 219 | But just for the sake of the example 220 | 221 | 56 222 | 00:03:01,140 --> 00:03:04,300 223 | well, we've already extracted the sentiment from the review 224 | 225 | 57 226 | 00:03:04,300 --> 00:03:08,500 227 | So here we have a response to the customer 228 | 229 | 58 230 | 00:03:08,500 --> 00:03:13,580 231 | It addresses details that the customer mentioned in their review 232 | 233 | 59 234 | 00:03:13,580 --> 00:03:16,620 235 | and as we instructed 236 | 237 | 60 238 | 00:03:16,620 --> 00:03:19,100 239 | suggest that they reach out to customer service 240 | 241 | 61 242 | 00:03:19,100 --> 00:03:23,100 243 | because this is just an AI customer service agent 244 | 245 | 62 246 | 00:03:23,740 --> 00:03:30,420 247 | Next, we're going to use a parameter of the language model called temperature 248 | 249 | 63 250 | 00:03:30,420 --> 00:03:36,460 251 | that will allow us to change the variety of the model's responses 252 | 253 | 64 254 | 00:03:36,460 --> 00:03:38,900 255 | So you can think of temperature as 256 | 257 | 65 258 | 00:03:38,900 --> 00:03:43,380 259 | the degree of exploration or randomness of the model 260 | 261 | 66 262 | 00:03:43,380 --> 00:03:46,060 263 | So for this particular phrase 264 | 265 | 67 266 | 00:03:46,060 --> 00:03:47,500 267 | my favorite food is 268 | 269 | 68 270 | 00:03:47,500 --> 00:03:51,820 271 | the most likely next word that the model predicts is pizza 272 | 273 | 69 274 | 00:03:51,820 --> 00:03:53,580 275 | and the next two most likely 276 | 277 | 70 278 | 00:03:53,580 --> 00:03:55,900 279 | it suggests are sushi and tacos 280 | 281 | 71 282 | 00:03:55,900 --> 00:03:57,660 283 | So at a temperature of zero 284 | 285 | 72 286 | 00:03:57,660 --> 00:04:00,060 287 | the model will always choose the most likely next word 288 | 289 | 73 290 | 00:04:00,060 --> 00:04:01,460 291 | which in this case is pizza 292 | 293 | 74 294 | 00:04:01,460 --> 00:04:03,260 295 | and at a higher temperature 296 | 297 | 75 298 | 00:04:03,260 --> 00:04:07,100 299 | it will also choose one of the less likely words 300 | 301 | 76 302 | 00:04:07,100 --> 00:04:09,340 303 | and at an even higher temperature 304 | 305 | 77 306 | 00:04:09,340 --> 00:04:10,700 307 | it might even choose tacos 308 | 309 | 78 310 | 00:04:10,700 --> 00:04:15,380 311 | which only has a five percent chance of being chosen 312 | 313 | 79 314 | 00:04:15,380 --> 00:04:21,660 315 | You can imagine that as the model continues this final response 316 | 317 | 80 318 | 00:04:21,660 --> 00:04:23,300 319 | so my favorite food is pizza 320 | 321 | 81 322 | 00:04:23,300 --> 00:04:25,420 323 | and it continues to generate more words 324 | 325 | 82 326 | 00:04:25,420 --> 00:04:30,580 327 | this response will diverge from the first response 328 | 329 | 83 330 | 00:04:30,580 --> 00:04:32,340 331 | which is my favorite food is tacos 332 | 333 | 84 334 | 00:04:32,340 --> 00:04:34,100 335 | So as the model continues 336 | 337 | 85 338 | 00:04:34,100 --> 00:04:37,060 339 | these two responses will become more and more different 340 | 341 | 86 342 | 00:04:37,060 --> 00:04:40,500 343 | In general, when building applications where you want 344 | 345 | 87 346 | 00:04:40,500 --> 00:04:43,420 347 | a predictable response 348 | 349 | 88 350 | 00:04:43,420 --> 00:04:45,420 351 | I would recommend using temperature zero 352 | 353 | 89 354 | 00:04:45,420 --> 00:04:46,780 355 | Throughout all of these videos 356 | 357 | 90 358 | 00:04:46,780 --> 00:04:48,900 359 | we've been using temperature zero 360 | 361 | 91 362 | 00:04:48,900 --> 00:04:51,380 363 | and I think that if you're trying to build a system that 364 | 365 | 92 366 | 00:04:51,380 --> 00:04:54,660 367 | is reliable and predictable, you should go with this 368 | 369 | 93 370 | 00:04:54,660 --> 00:04:58,780 371 | If you're trying to use the model in a more creative way 372 | 373 | 94 374 | 00:04:58,780 --> 00:05:04,220 375 | where you might want a wider variety of different outputs 376 | 377 | 95 378 | 00:05:04,220 --> 00:05:06,860 379 | you might want to use a higher temperature 380 | 381 | 96 382 | 00:05:06,860 --> 00:05:11,620 383 | So now let's take this same prompt that we just used 384 | 385 | 97 386 | 00:05:11,620 --> 00:05:14,220 387 | and let's try generating an e-mail 388 | 389 | 98 390 | 00:05:14,220 --> 00:05:16,140 391 | but let's use a higher temperature 392 | 393 | 99 394 | 00:05:16,140 --> 00:05:21,540 395 | So in our get_completion function that we've been using throughout the videos 396 | 397 | 100 398 | 00:05:21,540 --> 00:05:25,100 399 | we have specified a model and then also a temperature 400 | 401 | 101 402 | 00:05:25,100 --> 00:05:26,540 403 | but we've set them to default 404 | 405 | 102 406 | 00:05:26,540 --> 00:05:29,420 407 | So now let's try varying the temperature 408 | 409 | 103 410 | 00:05:30,140 --> 00:05:32,860 411 | So we use the prompt 412 | 413 | 104 414 | 00:05:32,860 --> 00:05:39,340 415 | and then let's try temperature=0.7 416 | 417 | 105 418 | 00:05:43,620 --> 00:05:45,340 419 | And so with temperature=0 420 | 421 | 106 422 | 00:05:45,340 --> 00:05:47,220 423 | every time you execute the same prompt 424 | 425 | 107 426 | 00:05:47,220 --> 00:05:50,580 427 | you should expect the same completion 428 | 429 | 108 430 | 00:05:50,580 --> 00:05:52,580 431 | whereas with temperature 0.7 432 | 433 | 109 434 | 00:05:52,580 --> 00:05:55,060 435 | you'll get a different output every time 436 | 437 | 110 438 | 00:05:55,060 --> 00:05:58,100 439 | So here we have our e-mail 440 | 441 | 111 442 | 00:05:58,100 --> 00:06:00,980 443 | and as you can see, it's different to the e-mail 444 | 445 | 112 446 | 00:06:00,980 --> 00:06:02,900 447 | that we received previously 448 | 449 | 113 450 | 00:06:02,900 --> 00:06:08,060 451 | And let's just execute it again to show that we'll get a different e-mail again 452 | 453 | 114 454 | 00:06:08,060 --> 00:06:11,340 455 | And here we have another different e-mail 456 | 457 | 115 458 | 00:06:11,340 --> 00:06:16,780 459 | So I recommend that you play around with temperature yourself 460 | 461 | 116 462 | 00:06:16,780 --> 00:06:20,820 463 | Maybe you could pause the video now and try this prompt 464 | 465 | 117 466 | 00:06:20,820 --> 00:06:22,620 467 | with a variety of different temperatures 468 | 469 | 118 470 | 00:06:22,620 --> 00:06:25,100 471 | just to see how the outputs vary 472 | 473 | 119 474 | 00:06:25,100 --> 00:06:28,500 475 | So to summarize, at higher temperatures 476 | 477 | 120 478 | 00:06:28,500 --> 00:06:31,700 479 | the outputs from the model are kind of more random 480 | 481 | 121 482 | 00:06:31,700 --> 00:06:34,500 483 | You can almost think of it as that at higher temperatures 484 | 485 | 122 486 | 00:06:34,500 --> 00:06:36,500 487 | the assistant is more distractible 488 | 489 | 123 490 | 00:06:36,500 --> 00:06:38,220 491 | but maybe more creative 492 | 493 | 124 494 | 00:06:39,620 --> 00:06:42,260 495 | In the next video, we're going to talk more about 496 | 497 | 125 498 | 00:06:42,260 --> 00:06:44,540 499 | the chat completions endpoint format 500 | 501 | 126 502 | 00:06:44,540 --> 00:07:00,060 503 | and how you can create a custom chatbot using this format 504 | 505 | -------------------------------------------------------------------------------- /chinese_subtitles/prompt_eng_06_transforming_chinese.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:09,840 3 | 大语言模型非常善于将其输入转换为不同的格式 4 | 5 | 2 6 | 00:00:09,840 --> 00:00:14,100 7 | 例如，输入一种语言的文本并将其转换或 8 | 9 | 3 10 | 00:00:14,100 --> 00:00:19,300 11 | 把它翻译成不同的语言，或者帮助进行拼写和语法的修正 12 | 13 | 4 14 | 00:00:19,300 --> 00:00:21,780 15 | 因此，将一段可能不符合标准的语法文本作为输入 16 | 17 | 5 18 | 00:00:21,780 --> 00:00:25,020 19 | 并让它帮助你纠正拼写和语法 20 | 21 | 6 22 | 00:00:25,020 --> 00:00:28,380 23 | 或甚至转换格式，如 24 | 25 | 7 26 | 00:00:28,380 --> 00:00:31,100 27 | 输入HTML和输出JSON 28 | 29 | 8 30 | 00:00:31,100 --> 00:00:35,140 31 | 所以，我以前写的一些应用程序有点痛苦，用的是 32 | 33 | 9 34 | 00:00:35,140 --> 00:00:38,100 35 | 一堆正则表达式的应用 36 | 37 | 10 38 | 00:00:38,100 --> 00:00:41,800 39 | 现在用一个大语言模型和一些提示语就可以更简单地实现 40 | 41 | 11 42 | 00:00:41,800 --> 00:00:46,500 43 | 是的，我现在用ChatGPT来校对我写的几乎所有东西 44 | 45 | 12 46 | 00:00:46,500 --> 00:00:50,340 47 | 所以我很高兴现在向你展示Notebook中的一些更多的例子 48 | 49 | 13 50 | 00:00:50,340 --> 00:00:55,900 51 | 所以，首先，我们将导入openai，同时 52 | 53 | 14 54 | 00:00:55,900 --> 00:01:01,180 55 | 使用我们在整个视频中一直使用的get_completion辅助函数 56 | 57 | 15 58 | 00:01:01,180 --> 00:01:04,540 59 | 我们要做的第一件事是一个翻译任务 60 | 61 | 16 62 | 00:01:04,540 --> 00:01:09,940 63 | 因此，大语言模型是在许多来源的大量文本上训练出来的 64 | 65 | 17 66 | 00:01:09,940 --> 00:01:11,100 67 | 其中有很多是互联网 68 | 69 | 18 70 | 00:01:11,100 --> 00:01:13,600 71 | 这当然是用许多不同的语言 72 | 73 | 19 74 | 00:01:13,600 --> 00:01:17,580 75 | 因此，这使模型具有翻译能力 76 | 77 | 20 78 | 00:01:17,580 --> 00:01:22,380 79 | 这些模型知道数百种语言的不同程度的熟练程度 80 | 81 | 21 82 | 00:01:22,380 --> 00:01:26,540 83 | 所以我们将通过一些例子来说明如何使用这种能力 84 | 85 | 22 86 | 00:01:26,540 --> 00:01:29,580 87 | 所以，让我们从简单的东西开始 88 | 89 | 23 90 | 00:01:29,580 --> 00:01:31,540 91 | 所以在这第一个例子中 92 | 93 | 24 94 | 00:01:31,540 --> 00:01:35,700 95 | 提示是将以下英文文本翻译成西班牙语 96 | 97 | 25 98 | 00:01:35,700 --> 00:01:38,260 99 | 你好，我想订购一台搅拌机 100 | 101 | 26 102 | 00:01:38,260 --> 00:01:40,620 103 | 响应是：Hola 104 | 105 | 27 106 | 00:01:40,620 --> 00:01:45,580 107 | me gustaría ordenar una licuadora. 108 | 109 | 28 110 | 00:01:45,580 --> 00:01:48,660 111 | 我对所有讲西班牙语的人感到非常抱歉 112 | 113 | 29 114 | 00:01:48,660 --> 00:01:53,620 115 | 我从来没有学过西班牙语，很不幸，你肯定能看出来 116 | 117 | 30 118 | 00:01:54,500 --> 00:01:57,220 119 | 让我们尝试另一个例子 120 | 121 | 31 122 | 00:01:57,220 --> 00:02:00,460 123 | 所以在这个例子中 124 | 125 | 32 126 | 00:02:00,460 --> 00:02:04,220 127 | 提示是，告诉我这是什么语言 128 | 129 | 33 130 | 00:02:04,220 --> 00:02:06,980 131 | 那么这是用法语说的 132 | 133 | 34 134 | 00:02:06,980 --> 00:02:09,700 135 | combien coûte le lampadaire? 136 | 137 | 35 138 | 00:02:09,700 --> 00:02:12,020 139 | 所以我们来运行这个 140 | 141 | 36 142 | 00:02:12,820 --> 00:02:17,340 143 | 该模型已经识别出这是法语 144 | 145 | 37 146 | 00:02:17,340 --> 00:02:22,220 147 | 该模型还可以一次做多个翻译 148 | 149 | 38 150 | 00:02:22,220 --> 00:02:23,900 151 | 所以在这个例子中 152 | 153 | 39 154 | 00:02:23,900 --> 00:02:27,740 155 | 让我们说把下面的文字翻译成法语和西班牙语 156 | 157 | 40 158 | 00:02:27,740 --> 00:02:29,940 159 | 你知道吗，让我们再加一个 160 | 161 | 41 162 | 00:02:29,940 --> 00:02:34,060 163 | 英国海盗 164 | 165 | 42 166 | 00:02:34,380 --> 00:02:38,500 167 | 这段文字是，我想订购一个篮球 168 | 169 | 43 170 | 00:02:40,460 --> 00:02:43,180 171 | 所以这里我们有法语 172 | 173 | 44 174 | 00:02:43,180 --> 00:02:47,300 175 | 西班牙文，和英国海盗 176 | 177 | 45 178 | 00:02:47,300 --> 00:02:49,300 179 | 所以在某些语言中 180 | 181 | 46 182 | 00:02:49,300 --> 00:02:51,340 183 | 翻译可以改变，取决于 184 | 185 | 47 186 | 00:02:51,340 --> 00:02:53,660 187 | 说话人与听话人的关系 188 | 189 | 48 190 | 00:02:53,660 --> 00:02:56,860 191 | 你也可以向语言模型解释这一点 192 | 193 | 49 194 | 00:02:56,860 --> 00:03:00,380 195 | 这样它就能进行相应的翻译 196 | 197 | 50 198 | 00:03:00,380 --> 00:03:02,540 199 | 所以在这个例子中 200 | 201 | 51 202 | 00:03:02,540 --> 00:03:07,220 203 | 我们说把下面的文字以正式和非正式的形式翻译成西班牙语 204 | 205 | 52 206 | 00:03:07,220 --> 00:03:09,060 207 | 你想订购一个枕头吗？ 208 | 209 | 53 210 | 00:03:09,060 --> 00:03:11,260 211 | 还注意到，我们在这里使用了 212 | 213 | 54 214 | 00:03:11,260 --> 00:03:14,300 215 | 不同的分隔符，而不是这些双引号 216 | 217 | 55 218 | 00:03:14,300 --> 00:03:18,260 219 | 其实这并不重要，只要它是一个清晰的分隔 220 | 221 | 56 222 | 00:03:18,260 --> 00:03:22,700 223 | 所以这里我们有正式和非正式 224 | 225 | 57 226 | 00:03:22,700 --> 00:03:25,580 227 | 因此，正式的是当你和某人说话时，他是 228 | 229 | 58 230 | 00:03:25,580 --> 00:03:28,980 231 | 也许比你高级的人，或者你是在一个专业的情况下 232 | 233 | 59 234 | 00:03:28,980 --> 00:03:30,340 235 | 这时你要使用正式语气 236 | 237 | 60 238 | 00:03:30,340 --> 00:03:34,540 239 | 然后非正式的是当你和一群朋友说话时 240 | 241 | 61 242 | 00:03:34,540 --> 00:03:36,300 243 | 我实际上不会说西班牙语 244 | 245 | 62 246 | 00:03:36,300 --> 00:03:39,500 247 | 但我父亲会，他说这是正确的 248 | 249 | 63 250 | 00:03:39,500 --> 00:03:42,020 251 | 因此，对于 252 | 253 | 64 254 | 00:03:42,020 --> 00:03:47,900 255 | 我们要假装自己是一家跨国电子商务公司的负责人 256 | 257 | 65 258 | 00:03:47,900 --> 00:03:51,780 259 | 用户的信息将使用所有不同的语言 260 | 261 | 66 262 | 00:03:51,780 --> 00:03:57,340 263 | 因此，用户将用各种不同的语言告诉我们关于IT的问题 264 | 265 | 67 266 | 00:03:57,340 --> 00:04:00,300 267 | 所以我们需要一个通用的翻译器 268 | 269 | 68 270 | 00:04:00,300 --> 00:04:06,980 271 | 因此，首先，我们将粘贴一个各种不同语言的用户信息的列表 272 | 273 | 69 274 | 00:04:06,980 --> 00:04:15,660 275 | 现在，我们将循环浏览这些用户信息中的每一条 276 | 277 | 70 278 | 00:04:15,660 --> 00:04:19,900 279 | 因此，对于用户信息中的问题 280 | 281 | 71 282 | 00:04:19,980 --> 00:04:25,220 283 | 然后我将复制稍长的代码块 284 | 285 | 72 286 | 00:04:25,220 --> 00:04:28,740 287 | 所以我们要做的第一件事是要求模型 288 | 289 | 73 290 | 00:04:28,740 --> 00:04:32,420 291 | 告诉我们这个问题是什么语言的 292 | 293 | 74 294 | 00:04:32,420 --> 00:04:34,260 295 | 所以这里是提示 296 | 297 | 75 298 | 00:04:34,260 --> 00:04:40,380 299 | 然后，我们将打印出原始信息的语言和问题 300 | 301 | 76 302 | 00:04:40,380 --> 00:04:45,100 303 | 然后我们将要求模型将其翻译成英语和韩语 304 | 305 | 77 306 | 00:04:45,100 --> 00:04:51,860 307 | 所以让我们运行这个。所以，原始信息是法语 308 | 309 | 78 310 | 00:04:55,740 --> 00:04:58,300 311 | 所以我们有各种语言 312 | 313 | 79 314 | 00:04:58,300 --> 00:05:02,940 315 | 然后模型将它们翻译成英语和韩语 316 | 317 | 80 318 | 00:05:02,940 --> 00:05:06,180 319 | 你可以看到这里，所以模型说 320 | 321 | 81 322 | 00:05:06,180 --> 00:05:07,700 323 | 这是法语 324 | 325 | 82 326 | 00:05:07,700 --> 00:05:13,380 327 | 所以这是因为这个提示的反应将是，这是法语 328 | 329 | 83 330 | 00:05:13,380 --> 00:05:16,180 331 | 你可以尝试将这个提示编辑成这样的内容 332 | 333 | 84 334 | 00:05:16,180 --> 00:05:17,540 335 | 告诉我这是什么语言 336 | 337 | 85 338 | 00:05:17,540 --> 00:05:22,220 339 | 只用一个词来回答或不用句子 340 | 341 | 86 342 | 00:05:22,220 --> 00:05:25,340 343 | 如果你希望这只是一个词的话 344 | 345 | 87 346 | 00:05:25,340 --> 00:05:28,580 347 | 或者你可以要求它以JSON格式或类似的东西 348 | 349 | 88 350 | 00:05:28,580 --> 00:05:33,420 351 | 这可能会鼓励它不要使用整个句子 352 | 353 | 89 354 | 00:05:33,420 --> 00:05:38,620 355 | 因此，令人惊讶的是，你刚刚建立了一个通用翻译器 356 | 357 | 90 358 | 00:05:38,620 --> 00:05:45,180 359 | 也可以随时暂停视频，在这里添加任何你想尝试的其他语言 360 | 361 | 91 362 | 00:05:45,180 --> 00:05:47,340 363 | 也许是你自己说的语言 364 | 365 | 92 366 | 00:05:47,340 --> 00:05:49,860 367 | 看看这个模型是怎么做的 368 | 369 | 93 370 | 00:05:49,860 --> 00:05:55,060 371 | 因此，我们接下来要深入研究的是语气转换 372 | 373 | 94 374 | 00:05:55,060 --> 00:05:58,300 375 | 写作可以根据目标受众而有所不同 376 | 377 | 95 378 | 00:05:58,300 --> 00:06:03,140 379 | 我给同事或教授写电子邮件的方式 380 | 381 | 96 382 | 00:06:03,140 --> 00:06:06,820 383 | 显然会与我给我弟弟发短信的方式完全不同 384 | 385 | 97 386 | 00:06:06,820 --> 00:06:10,980 387 | 所以ChatGPT实际上也可以帮助产生不同的语气 388 | 389 | 98 390 | 00:06:10,980 --> 00:06:14,260 391 | 因此，让我们看一些例子 392 | 393 | 99 394 | 00:06:14,260 --> 00:06:16,620 395 | 在这第一个例子中 396 | 397 | 100 398 | 00:06:16,620 --> 00:06:20,780 399 | 提示是，将以下内容从俚语翻译成商务信函 400 | 401 | 101 402 | 00:06:20,780 --> 00:06:22,100 403 | 老兄，这是乔 404 | 405 | 102 406 | 00:06:22,100 --> 00:06:25,180 407 | 检查一下这盏落地灯的规格 408 | 409 | 103 410 | 00:06:25,180 --> 00:06:29,460 411 | 因此，让我们来执行这个 412 | 413 | 104 414 | 00:06:30,460 --> 00:06:35,500 415 | 正如你所看到的，我们有一封更正式的商业信函 416 | 417 | 105 418 | 00:06:35,500 --> 00:06:39,660 419 | 有一个关于落地灯规格的建议 420 | 421 | 106 422 | 00:06:39,660 --> 00:06:44,580 423 | 我们要做的下一件事是在不同的格式之间进行转换 424 | 425 | 107 426 | 00:06:44,580 --> 00:06:50,260 427 | ChatGPT非常善于在不同的格式之间进行转换，如JSON到HTML 428 | 429 | 108 430 | 00:06:50,260 --> 00:06:54,300 431 | XML，各种各样的东西，markdown 432 | 433 | 109 434 | 00:06:54,300 --> 00:06:58,660 435 | 所以在提示中，我们将描述输入和输出的格式 436 | 437 | 110 438 | 00:06:58,660 --> 00:07:01,900 439 | 所以这里有一个例子 440 | 441 | 111 442 | 00:07:01,900 --> 00:07:05,100 443 | 所以我们有这样一个JSON 444 | 445 | 112 446 | 00:07:05,100 --> 00:07:10,340 447 | 它包含一个餐厅员工的名单，包括他们的名字和电子邮件 448 | 449 | 113 450 | 00:07:10,340 --> 00:07:12,580 451 | 然后在提示中 452 | 453 | 114 454 | 00:07:12,580 --> 00:07:17,940 455 | 我们将要求模型将其从JSON翻译成HTML 456 | 457 | 115 458 | 00:07:17,940 --> 00:07:22,100 459 | 所以提示是，把下面的Python字典从JSON翻译成 460 | 461 | 116 462 | 00:07:22,100 --> 00:07:26,260 463 | HTML，包括列头和标题 464 | 465 | 117 466 | 00:07:28,220 --> 00:07:35,380 467 | 然后我们将从模型中得到响应并打印出来 468 | 469 | 118 470 | 00:07:35,380 --> 00:07:44,140 471 | 所以这里我们有一些HTML显示所有的雇员姓名和电子邮件 472 | 473 | 119 474 | 00:07:44,140 --> 00:07:49,220 475 | 那么现在让我们来看看我们是否可以作为 476 | 477 | 120 478 | 00:07:49,220 --> 00:07:54,460 479 | 所以我们要使用这个Python库中的这个显示函数 480 | 481 | 121 482 | 00:07:54,460 --> 00:07:58,860 483 | 显示 HTML 响应 484 | 485 | 122 486 | 00:08:00,100 --> 00:08:06,340 487 | 在这里你可以看到这是一个正确格式化的HTML表格 488 | 489 | 123 490 | 00:08:06,420 --> 00:08:13,100 491 | 我们要做的下一个转换任务是拼写检查和语法检查 492 | 493 | 124 494 | 00:08:13,100 --> 00:08:16,740 495 | 这是ChatGPT的一个非常流行的用途 496 | 497 | 125 498 | 00:08:16,740 --> 00:08:18,780 499 | 我强烈建议这样做 500 | 501 | 126 502 | 00:08:18,780 --> 00:08:20,060 503 | 我一直在做这个 504 | 505 | 127 506 | 00:08:20,060 --> 00:08:23,380 507 | 当你用非母语工作时，它尤其有用 508 | 509 | 128 510 | 00:08:23,380 --> 00:08:28,140 511 | 因此，这里有一些常见的语法和拼写问题的例子 512 | 513 | 129 514 | 00:08:28,140 --> 00:08:31,580 515 | 以及语言模型如何帮助解决这些问题 516 | 517 | 130 518 | 00:08:31,580 --> 00:08:40,500 519 | 所以我将粘贴一个有一些语法或拼写错误的句子列表 520 | 521 | 131 522 | 00:08:40,500 --> 00:08:46,420 523 | 然后，我们将循环浏览每一个句子 524 | 525 | 132 526 | 00:08:49,140 --> 00:08:50,740 527 | 然后 528 | 529 | 133 530 | 00:08:51,620 --> 00:08:56,620 531 | 让模型进行校对 532 | 533 | 134 534 | 00:08:56,620 --> 00:09:02,500 535 | 校对并进行修正。然后我们将使用一些分隔符 536 | 537 | 135 538 | 00:09:11,500 --> 00:09:15,500 539 | 然后我们将获取响应并将其打印出来，就像往常一样 540 | 541 | 136 542 | 00:09:35,260 --> 00:09:41,420 543 | 就这样，这个模型能够纠正所有这些语法错误 544 | 545 | 137 546 | 00:09:41,420 --> 00:09:44,180 547 | 我们可以使用我们之前讨论过的一些技术 548 | 549 | 138 550 | 00:09:44,180 --> 00:09:50,100 551 | 因此，为了改进提示，我们可以说校对和纠正以下文字 552 | 553 | 139 554 | 00:09:50,100 --> 00:09:54,580 555 | 并重写 556 | 557 | 140 558 | 00:09:54,580 --> 00:09:55,580 559 | 哦 560 | 561 | 141 562 | 00:09:55,580 --> 00:10:05,580 563 | 并重写整个 564 | 565 | 142 566 | 00:10:05,580 --> 00:10:10,580 567 | 并重写 568 | 569 | 143 570 | 00:10:10,580 --> 00:10:16,580 571 | 改正后的版本 572 | 573 | 144 574 | 00:10:16,580 --> 00:10:25,700 575 | 如果你没有发现任何错误，就说没有发现错误 576 | 577 | 145 578 | 00:10:25,700 --> 00:10:29,700 579 | 让我们来试试这个 580 | 581 | 146 582 | 00:10:29,700 --> 00:10:37,700 583 | 因此，这种方式我们能够，哦，他们在这里仍然使用引号 584 | 585 | 147 586 | 00:10:37,700 --> 00:10:42,940 587 | 但你可以想象，你能够找到一种方法，用一点点的 588 | 589 | 148 590 | 00:10:42,940 --> 00:10:45,780 591 | 迭代提示开发 592 | 593 | 149 594 | 00:10:45,780 --> 00:10:49,260 595 | 找到一种更可靠的提示方式，每一次都能正常工作 596 | 597 | 150 598 | 00:10:49,260 --> 00:10:51,700 599 | 因此，现在我们将实现另一个例子 600 | 601 | 151 602 | 00:10:51,700 --> 00:10:56,040 603 | 在你把你的文本发布到公共论坛之前，检查它总是有用的 604 | 605 | 152 606 | 00:10:56,040 --> 00:10:59,100 607 | 因此，我们将通过一个检查评论的例子 608 | 609 | 153 610 | 00:10:59,100 --> 00:11:01,980 611 | 这里有一个关于毛绒熊猫的评论 612 | 613 | 154 614 | 00:11:01,980 --> 00:11:06,740 615 | 因此，我们将要求模型校对和纠正评论 616 | 617 | 155 618 | 00:11:06,740 --> 00:11:11,100 619 | 很好。所以我们有了这个更正的版本 620 | 621 | 156 622 | 00:11:11,100 --> 00:11:14,860 623 | 我们可以做的一件很酷的事情是找到 624 | 625 | 157 626 | 00:11:14,860 --> 00:11:19,180 627 | 我们的原始评论和模型的输出之间的差异 628 | 629 | 158 630 | 00:11:19,180 --> 00:11:24,380 631 | 所以我们要基于这个Redlines Python包来做这个 632 | 633 | 159 634 | 00:11:24,380 --> 00:11:26,800 635 | 我们将得到以下的差异 636 | 637 | 160 638 | 00:11:26,800 --> 00:11:32,460 639 | 我们评论的原文和模型的输出之间的差异 640 | 641 | 161 642 | 00:11:32,460 --> 00:11:36,740 643 | 然后显示这个 644 | 645 | 162 646 | 00:11:36,740 --> 00:11:40,340 647 | 所以在这里你可以看到原始评论和模型输出之间的差异 648 | 649 | 163 650 | 00:11:40,340 --> 00:11:44,940 651 | 和模型输出之间的差异，以及那些已经被纠正的东西 652 | 653 | 164 654 | 00:11:44,940 --> 00:11:49,380 655 | 因此，我们使用的提示是校对和纠正这个评论 656 | 657 | 165 658 | 00:11:49,380 --> 00:11:52,820 659 | 但你也可以做一些更戏剧性的改变 660 | 661 | 166 662 | 00:11:52,820 --> 00:11:55,700 663 | 语气的变化，以及诸如此类的事情 664 | 665 | 167 666 | 00:11:55,700 --> 00:12:00,340 667 | 因此，让我们再试一试 668 | 669 | 168 670 | 00:12:00,340 --> 00:12:04,500 671 | 所以在这个提示中，我们将要求模特校对和纠正 672 | 673 | 169 674 | 00:12:04,500 --> 00:12:08,700 675 | 同样的评论，但也要使其更有说服力，并确保其 676 | 677 | 170 678 | 00:12:08,700 --> 00:12:11,860 679 | 遵循APA风格，并针对高级读者 680 | 681 | 171 682 | 00:12:11,860 --> 00:12:14,860 683 | 我们还将要求以Markdown格式输出结果 684 | 685 | 172 686 | 00:12:14,860 --> 00:12:18,780 687 | 因此，我们在这里使用原始审查的相同文本 688 | 689 | 173 690 | 00:12:18,780 --> 00:12:23,260 691 | 因此，让我们来执行这个 692 | 693 | 174 694 | 00:12:23,260 --> 00:12:30,380 695 | 在这里，我们有一个扩展的APA风格的毛绒熊猫的评论 696 | 697 | 175 698 | 00:12:30,380 --> 00:12:32,300 699 | 因此，这是它的转换视频 700 | 701 | 176 702 | 00:12:32,300 --> 00:12:35,700 703 | 接下来，我们将使用扩写功能，采取较短的提示 704 | 705 | 177 706 | 00:12:35,700 --> 00:12:54,180 707 | 并从语言模型中产生一个更长、更自由的反应 708 | 709 | -------------------------------------------------------------------------------- /course_notebooks/l5-inferring.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "attachments": {}, 5 | "cell_type": "markdown", 6 | "metadata": {}, 7 | "source": [ 8 | "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/GitHubDaily/ChatGPT-Prompt-Engineering-for-Developers-in-Chinese/blob/master/assets/5-Inferring/l5-inferring.ipynb)\n", 9 | "# Inferring\n", 10 | "In this lesson, you will infer sentiment and topics from product reviews and news articles.\n", 11 | "\n", 12 | "## Setup" 13 | ] 14 | }, 15 | { 16 | "cell_type": "code", 17 | "execution_count": null, 18 | "metadata": {}, 19 | "outputs": [], 20 | "source": [ 21 | "# Basic congfig \n", 22 | "# Install basic package and set key\n", 23 | "!pip install openai\n", 24 | "!export OPENAI_API_KEY='sk-...'" 25 | ] 26 | }, 27 | { 28 | "cell_type": "code", 29 | "execution_count": null, 30 | "metadata": {}, 31 | "outputs": [], 32 | "source": [ 33 | "import openai\n", 34 | "import os\n", 35 | "\n", 36 | "from dotenv import load_dotenv, find_dotenv\n", 37 | "_ = load_dotenv(find_dotenv()) # read local .env file\n", 38 | "\n", 39 | "openai.api_key = os.getenv('OPENAI_API_KEY')" 40 | ] 41 | }, 42 | { 43 | "cell_type": "code", 44 | "execution_count": null, 45 | "metadata": {}, 46 | "outputs": [], 47 | "source": [ 48 | "def get_completion(prompt, model=\"gpt-3.5-turbo\"):\n", 49 | " messages = [{\"role\": \"user\", \"content\": prompt}]\n", 50 | " response = openai.ChatCompletion.create(\n", 51 | " model=model,\n", 52 | " messages=messages,\n", 53 | " temperature=0, # this is the degree of randomness of the model's output\n", 54 | " )\n", 55 | " return response.choices[0].message[\"content\"]" 56 | ] 57 | }, 58 | { 59 | "attachments": {}, 60 | "cell_type": "markdown", 61 | "metadata": {}, 62 | "source": [ 63 | "## Product review text" 64 | ] 65 | }, 66 | { 67 | "cell_type": "code", 68 | "execution_count": null, 69 | "metadata": {}, 70 | "outputs": [], 71 | "source": [ 72 | "lamp_review = \"\"\"\n", 73 | "Needed a nice lamp for my bedroom, and this one had \\\n", 74 | "additional storage and not too high of a price point. \\\n", 75 | "Got it fast. The string to our lamp broke during the \\\n", 76 | "transit and the company happily sent over a new one. \\\n", 77 | "Came within a few days as well. It was easy to put \\\n", 78 | "together. I had a missing part, so I contacted their \\\n", 79 | "support and they very quickly got me the missing piece! \\\n", 80 | "Lumina seems to me to be a great company that cares \\\n", 81 | "about their customers and products!!\n", 82 | "\"\"\"" 83 | ] 84 | }, 85 | { 86 | "attachments": {}, 87 | "cell_type": "markdown", 88 | "metadata": {}, 89 | "source": [ 90 | "## Sentiment (positive/negative)\n" 91 | ] 92 | }, 93 | { 94 | "cell_type": "code", 95 | "execution_count": null, 96 | "metadata": {}, 97 | "outputs": [], 98 | "source": [ 99 | "prompt = f\"\"\"\n", 100 | "What is the sentiment of the following product review, \n", 101 | "which is delimited with triple backticks?\n", 102 | "\n", 103 | "Review text: '''{lamp_review}'''\n", 104 | "\"\"\"\n", 105 | "response = get_completion(prompt)\n", 106 | "print(response)" 107 | ] 108 | }, 109 | { 110 | "cell_type": "code", 111 | "execution_count": null, 112 | "metadata": {}, 113 | "outputs": [], 114 | "source": [ 115 | "prompt = f\"\"\"\n", 116 | "What is the sentiment of the following product review, \n", 117 | "which is delimited with triple backticks?\n", 118 | "\n", 119 | "Give your answer as a single word, either \"positive\" \\\n", 120 | "or \"negative\".\n", 121 | "\n", 122 | "Review text: '''{lamp_review}'''\n", 123 | "\"\"\"\n", 124 | "response = get_completion(prompt)\n", 125 | "print(response)" 126 | ] 127 | }, 128 | { 129 | "attachments": {}, 130 | "cell_type": "markdown", 131 | "metadata": {}, 132 | "source": [ 133 | "## Identify types of emotions" 134 | ] 135 | }, 136 | { 137 | "cell_type": "code", 138 | "execution_count": null, 139 | "metadata": {}, 140 | "outputs": [], 141 | "source": [ 142 | "prompt = f\"\"\"\n", 143 | "Identify a list of emotions that the writer of the \\\n", 144 | "following review is expressing. Include no more than \\\n", 145 | "five items in the list. Format your answer as a list of \\\n", 146 | "lower-case words separated by commas.\n", 147 | "\n", 148 | "Review text: '''{lamp_review}'''\n", 149 | "\"\"\"\n", 150 | "response = get_completion(prompt)\n", 151 | "print(response)" 152 | ] 153 | }, 154 | { 155 | "attachments": {}, 156 | "cell_type": "markdown", 157 | "metadata": {}, 158 | "source": [ 159 | "## Identify anger" 160 | ] 161 | }, 162 | { 163 | "cell_type": "code", 164 | "execution_count": null, 165 | "metadata": {}, 166 | "outputs": [], 167 | "source": [ 168 | "prompt = f\"\"\"\n", 169 | "Is the writer of the following review expressing anger?\\\n", 170 | "The review is delimited with triple backticks. \\\n", 171 | "Give your answer as either yes or no.\n", 172 | "\n", 173 | "Review text: '''{lamp_review}'''\n", 174 | "\"\"\"\n", 175 | "response = get_completion(prompt)\n", 176 | "print(response)" 177 | ] 178 | }, 179 | { 180 | "attachments": {}, 181 | "cell_type": "markdown", 182 | "metadata": {}, 183 | "source": [ 184 | "## Extract product and company name from customer reviews" 185 | ] 186 | }, 187 | { 188 | "cell_type": "code", 189 | "execution_count": null, 190 | "metadata": {}, 191 | "outputs": [], 192 | "source": [ 193 | "prompt = f\"\"\"\n", 194 | "Identify the following items from the review text: \n", 195 | "- Item purchased by reviewer\n", 196 | "- Company that made the item\n", 197 | "\n", 198 | "The review is delimited with triple backticks. \\\n", 199 | "Format your response as a JSON object with \\\n", 200 | "\"Item\" and \"Brand\" as the keys. \n", 201 | "If the information isn't present, use \"unknown\" \\\n", 202 | "as the value.\n", 203 | "Make your response as short as possible.\n", 204 | " \n", 205 | "Review text: '''{lamp_review}'''\n", 206 | "\"\"\"\n", 207 | "response = get_completion(prompt)\n", 208 | "print(response)" 209 | ] 210 | }, 211 | { 212 | "attachments": {}, 213 | "cell_type": "markdown", 214 | "metadata": {}, 215 | "source": [ 216 | "## Doing multiple tasks at once" 217 | ] 218 | }, 219 | { 220 | "cell_type": "code", 221 | "execution_count": null, 222 | "metadata": {}, 223 | "outputs": [], 224 | "source": [ 225 | "prompt = f\"\"\"\n", 226 | "Identify the following items from the review text: \n", 227 | "- Sentiment (positive or negative)\n", 228 | "- Is the reviewer expressing anger? (true or false)\n", 229 | "- Item purchased by reviewer\n", 230 | "- Company that made the item\n", 231 | "\n", 232 | "The review is delimited with triple backticks. \\\n", 233 | "Format your response as a JSON object with \\\n", 234 | "\"Sentiment\", \"Anger\", \"Item\" and \"Brand\" as the keys.\n", 235 | "If the information isn't present, use \"unknown\" \\\n", 236 | "as the value.\n", 237 | "Make your response as short as possible.\n", 238 | "Format the Anger value as a boolean.\n", 239 | "\n", 240 | "Review text: '''{lamp_review}'''\n", 241 | "\"\"\"\n", 242 | "response = get_completion(prompt)\n", 243 | "print(response)" 244 | ] 245 | }, 246 | { 247 | "attachments": {}, 248 | "cell_type": "markdown", 249 | "metadata": {}, 250 | "source": [ 251 | "## Inferring topics" 252 | ] 253 | }, 254 | { 255 | "cell_type": "code", 256 | "execution_count": null, 257 | "metadata": {}, 258 | "outputs": [], 259 | "source": [ 260 | "story = \"\"\"\n", 261 | "In a recent survey conducted by the government, \n", 262 | "public sector employees were asked to rate their level \n", 263 | "of satisfaction with the department they work at. \n", 264 | "The results revealed that NASA was the most popular \n", 265 | "department with a satisfaction rating of 95%.\n", 266 | "\n", 267 | "One NASA employee, John Smith, commented on the findings, \n", 268 | "stating, \"I'm not surprised that NASA came out on top. \n", 269 | "It's a great place to work with amazing people and \n", 270 | "incredible opportunities. I'm proud to be a part of \n", 271 | "such an innovative organization.\"\n", 272 | "\n", 273 | "The results were also welcomed by NASA's management team, \n", 274 | "with Director Tom Johnson stating, \"We are thrilled to \n", 275 | "hear that our employees are satisfied with their work at NASA. \n", 276 | "We have a talented and dedicated team who work tirelessly \n", 277 | "to achieve our goals, and it's fantastic to see that their \n", 278 | "hard work is paying off.\"\n", 279 | "\n", 280 | "The survey also revealed that the \n", 281 | "Social Security Administration had the lowest satisfaction \n", 282 | "rating, with only 45% of employees indicating they were \n", 283 | "satisfied with their job. The government has pledged to \n", 284 | "address the concerns raised by employees in the survey and \n", 285 | "work towards improving job satisfaction across all departments.\n", 286 | "\"\"\"" 287 | ] 288 | }, 289 | { 290 | "attachments": {}, 291 | "cell_type": "markdown", 292 | "metadata": {}, 293 | "source": [ 294 | "## Infer 5 topics" 295 | ] 296 | }, 297 | { 298 | "cell_type": "code", 299 | "execution_count": null, 300 | "metadata": {}, 301 | "outputs": [], 302 | "source": [ 303 | "prompt = f\"\"\"\n", 304 | "Determine five topics that are being discussed in the \\\n", 305 | "following text, which is delimited by triple backticks.\n", 306 | "\n", 307 | "Make each item one or two words long. \n", 308 | "\n", 309 | "Format your response as a list of items separated by commas.\n", 310 | "\n", 311 | "Text sample: '''{story}'''\n", 312 | "\"\"\"\n", 313 | "response = get_completion(prompt)\n", 314 | "print(response)" 315 | ] 316 | }, 317 | { 318 | "cell_type": "code", 319 | "execution_count": null, 320 | "metadata": {}, 321 | "outputs": [], 322 | "source": [ 323 | "response.split(sep=',')" 324 | ] 325 | }, 326 | { 327 | "cell_type": "code", 328 | "execution_count": null, 329 | "metadata": {}, 330 | "outputs": [], 331 | "source": [ 332 | "topic_list = [\n", 333 | " \"nasa\", \"local government\", \"engineering\", \n", 334 | " \"employee satisfaction\", \"federal government\"\n", 335 | "]" 336 | ] 337 | }, 338 | { 339 | "attachments": {}, 340 | "cell_type": "markdown", 341 | "metadata": {}, 342 | "source": [ 343 | "## Make a news alert for certain topics" 344 | ] 345 | }, 346 | { 347 | "cell_type": "code", 348 | "execution_count": null, 349 | "metadata": {}, 350 | "outputs": [], 351 | "source": [ 352 | "prompt = f\"\"\"\n", 353 | "Determine whether each item in the following list of \\\n", 354 | "topics is a topic in the text below, which\n", 355 | "is delimited with triple backticks.\n", 356 | "\n", 357 | "Give your answer as list with 0 or 1 for each topic.\\\n", 358 | "\n", 359 | "List of topics: {\", \".join(topic_list)}\n", 360 | "\n", 361 | "Text sample: '''{story}'''\n", 362 | "\"\"\"\n", 363 | "response = get_completion(prompt)\n", 364 | "print(response)" 365 | ] 366 | }, 367 | { 368 | "cell_type": "code", 369 | "execution_count": null, 370 | "metadata": {}, 371 | "outputs": [], 372 | "source": [ 373 | "topic_dict = {i.split(': ')[0]: int(i.split(': ')[1]) for i in response.split(sep='\\n')}\n", 374 | "if topic_dict['nasa'] == 1:\n", 375 | " print(\"ALERT: New NASA story!\")" 376 | ] 377 | }, 378 | { 379 | "cell_type": "markdown", 380 | "metadata": {}, 381 | "source": [] 382 | } 383 | ], 384 | "metadata": { 385 | "kernelspec": { 386 | "display_name": "gpt_index", 387 | "language": "python", 388 | "name": "python3" 389 | }, 390 | "language_info": { 391 | "codemirror_mode": { 392 | "name": "ipython", 393 | "version": 3 394 | }, 395 | "file_extension": ".py", 396 | "mimetype": "text/x-python", 397 | "name": "python", 398 | "nbconvert_exporter": "python", 399 | "pygments_lexer": "ipython3", 400 | "version": "3.8.16" 401 | }, 402 | "orig_nbformat": 4 403 | }, 404 | "nbformat": 4, 405 | "nbformat_minor": 2 406 | } 407 | -------------------------------------------------------------------------------- /course_notebooks/l6-transforming.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "attachments": {}, 5 | "cell_type": "markdown", 6 | "metadata": {}, 7 | "source": [ 8 | "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/GitHubDaily/ChatGPT-Prompt-Engineering-for-Developers-in-Chinese/blob/master/assets/6-Transforming/l6-transforming.ipynb)\n", 9 | "# Transforming\n", 10 | "\n", 11 | "In this notebook, we will explore how to use Large Language Models for text transformation tasks such as language translation, spelling and grammar checking, tone adjustment, and format conversion.\n", 12 | "\n", 13 | "## Setup" 14 | ] 15 | }, 16 | { 17 | "cell_type": "code", 18 | "execution_count": null, 19 | "metadata": {}, 20 | "outputs": [], 21 | "source": [ 22 | "# Basic congfig \n", 23 | "# Install basic package and set key\n", 24 | "!pip install openai\n", 25 | "!export OPENAI_API_KEY='sk-...'" 26 | ] 27 | }, 28 | { 29 | "cell_type": "code", 30 | "execution_count": null, 31 | "metadata": {}, 32 | "outputs": [], 33 | "source": [ 34 | "import openai\n", 35 | "import os\n", 36 | "\n", 37 | "from dotenv import load_dotenv, find_dotenv\n", 38 | "_ = load_dotenv(find_dotenv()) # read local .env file\n", 39 | "\n", 40 | "openai.api_key = os.getenv('OPENAI_API_KEY')" 41 | ] 42 | }, 43 | { 44 | "cell_type": "code", 45 | "execution_count": null, 46 | "metadata": {}, 47 | "outputs": [], 48 | "source": [ 49 | "def get_completion(prompt, model=\"gpt-3.5-turbo\", temperature=0): \n", 50 | " messages = [{\"role\": \"user\", \"content\": prompt}]\n", 51 | " response = openai.ChatCompletion.create(\n", 52 | " model=model,\n", 53 | " messages=messages,\n", 54 | " temperature=temperature, \n", 55 | " )\n", 56 | " return response.choices[0].message[\"content\"]" 57 | ] 58 | }, 59 | { 60 | "attachments": {}, 61 | "cell_type": "markdown", 62 | "metadata": {}, 63 | "source": [ 64 | "## Translation\n", 65 | "\n", 66 | "ChatGPT is trained with sources in many languages. This gives the model the ability to do translation. Here are some examples of how to use this capability." 67 | ] 68 | }, 69 | { 70 | "cell_type": "code", 71 | "execution_count": null, 72 | "metadata": {}, 73 | "outputs": [], 74 | "source": [ 75 | "prompt = f\"\"\"\n", 76 | "Translate the following English text to Spanish: \\ \n", 77 | "```Hi, I would like to order a blender```\n", 78 | "\"\"\"\n", 79 | "response = get_completion(prompt)\n", 80 | "print(response)" 81 | ] 82 | }, 83 | { 84 | "cell_type": "code", 85 | "execution_count": null, 86 | "metadata": {}, 87 | "outputs": [], 88 | "source": [ 89 | "prompt = f\"\"\"\n", 90 | "Tell me which language this is: \n", 91 | "```Combien coûte le lampadaire?```\n", 92 | "\"\"\"\n", 93 | "response = get_completion(prompt)\n", 94 | "print(response)" 95 | ] 96 | }, 97 | { 98 | "cell_type": "code", 99 | "execution_count": null, 100 | "metadata": {}, 101 | "outputs": [], 102 | "source": [ 103 | "prompt = f\"\"\"\n", 104 | "Translate the following text to French and Spanish\n", 105 | "and English pirate: \\\n", 106 | "```I want to order a basketball```\n", 107 | "\"\"\"\n", 108 | "response = get_completion(prompt)\n", 109 | "print(response)" 110 | ] 111 | }, 112 | { 113 | "cell_type": "code", 114 | "execution_count": null, 115 | "metadata": {}, 116 | "outputs": [], 117 | "source": [ 118 | "prompt = f\"\"\"\n", 119 | "Translate the following text to Spanish in both the \\\n", 120 | "formal and informal forms: \n", 121 | "'Would you like to order a pillow?'\n", 122 | "\"\"\"\n", 123 | "response = get_completion(prompt)\n", 124 | "print(response)" 125 | ] 126 | }, 127 | { 128 | "attachments": {}, 129 | "cell_type": "markdown", 130 | "metadata": {}, 131 | "source": [ 132 | "### Universal Translator\n", 133 | "Imagine you are in charge of IT at a large multinational e-commerce company. Users are messaging you with IT issues in all their native languages. Your staff is from all over the world and speaks only their native languages. You need a universal translator!" 134 | ] 135 | }, 136 | { 137 | "cell_type": "code", 138 | "execution_count": null, 139 | "metadata": {}, 140 | "outputs": [], 141 | "source": [ 142 | "user_messages = [\n", 143 | " \"La performance du système est plus lente que d'habitude.\", # System performance is slower than normal \n", 144 | " \"Mi monitor tiene píxeles que no se iluminan.\", # My monitor has pixels that are not lighting\n", 145 | " \"Il mio mouse non funziona\", # My mouse is not working\n", 146 | " \"Mój klawisz Ctrl jest zepsuty\", # My keyboard has a broken control key\n", 147 | " \"我的屏幕在闪烁\" # My screen is flashing\n", 148 | "] " 149 | ] 150 | }, 151 | { 152 | "cell_type": "code", 153 | "execution_count": null, 154 | "metadata": {}, 155 | "outputs": [], 156 | "source": [ 157 | "for issue in user_messages:\n", 158 | " prompt = f\"Tell me what language this is: ```{issue}```\"\n", 159 | " lang = get_completion(prompt)\n", 160 | " print(f\"Original message ({lang}): {issue}\")\n", 161 | "\n", 162 | " prompt = f\"\"\"\n", 163 | " Translate the following text to English \\\n", 164 | " and Korean: ```{issue}```\n", 165 | " \"\"\"\n", 166 | " response = get_completion(prompt)\n", 167 | " print(response, \"\\n\")" 168 | ] 169 | }, 170 | { 171 | "attachments": {}, 172 | "cell_type": "markdown", 173 | "metadata": {}, 174 | "source": [ 175 | "## Try it yourself!\n", 176 | "Try some translations on your own!" 177 | ] 178 | }, 179 | { 180 | "cell_type": "code", 181 | "execution_count": null, 182 | "metadata": {}, 183 | "outputs": [], 184 | "source": [] 185 | }, 186 | { 187 | "attachments": {}, 188 | "cell_type": "markdown", 189 | "metadata": {}, 190 | "source": [ 191 | "## Tone Transformation\n", 192 | "Writing can vary based on the intended audience. ChatGPT can produce different tones.\n" 193 | ] 194 | }, 195 | { 196 | "cell_type": "code", 197 | "execution_count": null, 198 | "metadata": {}, 199 | "outputs": [], 200 | "source": [ 201 | "prompt = f\"\"\"\n", 202 | "Translate the following from slang to a business letter: \n", 203 | "'Dude, This is Joe, check out this spec on this standing lamp.'\n", 204 | "\"\"\"\n", 205 | "response = get_completion(prompt)\n", 206 | "print(response)" 207 | ] 208 | }, 209 | { 210 | "attachments": {}, 211 | "cell_type": "markdown", 212 | "metadata": {}, 213 | "source": [ 214 | "## Format Conversion\n", 215 | "ChatGPT can translate between formats. The prompt should describe the input and output formats." 216 | ] 217 | }, 218 | { 219 | "cell_type": "code", 220 | "execution_count": null, 221 | "metadata": {}, 222 | "outputs": [], 223 | "source": [ 224 | "data_json = { \"resturant employees\" :[ \n", 225 | " {\"name\":\"Shyam\", \"email\":\"shyamjaiswal@gmail.com\"},\n", 226 | " {\"name\":\"Bob\", \"email\":\"bob32@gmail.com\"},\n", 227 | " {\"name\":\"Jai\", \"email\":\"jai87@gmail.com\"}\n", 228 | "]}\n", 229 | "\n", 230 | "prompt = f\"\"\"\n", 231 | "Translate the following python dictionary from JSON to an HTML \\\n", 232 | "table with column headers and title: {data_json}\n", 233 | "\"\"\"\n", 234 | "response = get_completion(prompt)\n", 235 | "print(response)" 236 | ] 237 | }, 238 | { 239 | "cell_type": "code", 240 | "execution_count": null, 241 | "metadata": {}, 242 | "outputs": [], 243 | "source": [ 244 | "from IPython.display import display, Markdown, Latex, HTML, JSON\n", 245 | "display(HTML(response))" 246 | ] 247 | }, 248 | { 249 | "attachments": {}, 250 | "cell_type": "markdown", 251 | "metadata": {}, 252 | "source": [ 253 | "## Spellcheck/Grammar check.\n", 254 | "\n", 255 | "Here are some examples of common grammar and spelling problems and the LLM's response. \n", 256 | "\n", 257 | "To signal to the LLM that you want it to proofread your text, you instruct the model to 'proofread' or 'proofread and correct'." 258 | ] 259 | }, 260 | { 261 | "cell_type": "code", 262 | "execution_count": null, 263 | "metadata": {}, 264 | "outputs": [], 265 | "source": [ 266 | "text = [ \n", 267 | " \"The girl with the black and white puppies have a ball.\", # The girl has a ball.\n", 268 | " \"Yolanda has her notebook.\", # ok\n", 269 | " \"Its going to be a long day. Does the car need it’s oil changed?\", # Homonyms\n", 270 | " \"Their goes my freedom. There going to bring they’re suitcases.\", # Homonyms\n", 271 | " \"Your going to need you’re notebook.\", # Homonyms\n", 272 | " \"That medicine effects my ability to sleep. Have you heard of the butterfly affect?\", # Homonyms\n", 273 | " \"This phrase is to cherck chatGPT for speling abilitty\" # spelling\n", 274 | "]\n", 275 | "for t in text:\n", 276 | " prompt = f\"\"\"Proofread and correct the following text\n", 277 | " and rewrite the corrected version. If you don't find\n", 278 | " and errors, just say \"No errors found\". Don't use \n", 279 | " any punctuation around the text:\n", 280 | " ```{t}```\"\"\"\n", 281 | " response = get_completion(prompt)\n", 282 | " print(response)" 283 | ] 284 | }, 285 | { 286 | "cell_type": "code", 287 | "execution_count": null, 288 | "metadata": {}, 289 | "outputs": [], 290 | "source": [ 291 | "text = f\"\"\"\n", 292 | "Got this for my daughter for her birthday cuz she keeps taking \\\n", 293 | "mine from my room. Yes, adults also like pandas too. She takes \\\n", 294 | "it everywhere with her, and it's super soft and cute. One of the \\\n", 295 | "ears is a bit lower than the other, and I don't think that was \\\n", 296 | "designed to be asymmetrical. It's a bit small for what I paid for it \\\n", 297 | "though. I think there might be other options that are bigger for \\\n", 298 | "the same price. It arrived a day earlier than expected, so I got \\\n", 299 | "to play with it myself before I gave it to my daughter.\n", 300 | "\"\"\"\n", 301 | "prompt = f\"proofread and correct this review: ```{text}```\"\n", 302 | "response = get_completion(prompt)\n", 303 | "print(response)" 304 | ] 305 | }, 306 | { 307 | "cell_type": "code", 308 | "execution_count": null, 309 | "metadata": {}, 310 | "outputs": [], 311 | "source": [ 312 | "!pip install redlines\n", 313 | "from redlines import Redlines\n", 314 | "\n", 315 | "diff = Redlines(text,response)\n", 316 | "display(Markdown(diff.output_markdown))" 317 | ] 318 | }, 319 | { 320 | "cell_type": "code", 321 | "execution_count": null, 322 | "metadata": {}, 323 | "outputs": [], 324 | "source": [ 325 | "prompt = f\"\"\"\n", 326 | "proofread and correct this review. Make it more compelling. \n", 327 | "Ensure it follows APA style guide and targets an advanced reader. \n", 328 | "Output in markdown format.\n", 329 | "Text: ```{text}```\n", 330 | "\"\"\"\n", 331 | "response = get_completion(prompt)\n", 332 | "display(Markdown(response))" 333 | ] 334 | }, 335 | { 336 | "attachments": {}, 337 | "cell_type": "markdown", 338 | "metadata": {}, 339 | "source": [ 340 | "## Try it yourself!\n", 341 | "Try changing the instructions to form your own review." 342 | ] 343 | }, 344 | { 345 | "cell_type": "markdown", 346 | "metadata": {}, 347 | "source": [] 348 | }, 349 | { 350 | "cell_type": "code", 351 | "execution_count": null, 352 | "metadata": {}, 353 | "outputs": [], 354 | "source": [] 355 | }, 356 | { 357 | "attachments": {}, 358 | "cell_type": "markdown", 359 | "metadata": {}, 360 | "source": [ 361 | "Thanks to the following sites:\n", 362 | "\n", 363 | "https://writingprompts.com/bad-grammar-examples/" 364 | ] 365 | }, 366 | { 367 | "cell_type": "markdown", 368 | "metadata": {}, 369 | "source": [] 370 | } 371 | ], 372 | "metadata": { 373 | "kernelspec": { 374 | "display_name": "gpt_index", 375 | "language": "python", 376 | "name": "python3" 377 | }, 378 | "language_info": { 379 | "codemirror_mode": { 380 | "name": "ipython", 381 | "version": 3 382 | }, 383 | "file_extension": ".py", 384 | "mimetype": "text/x-python", 385 | "name": "python", 386 | "nbconvert_exporter": "python", 387 | "pygments_lexer": "ipython3", 388 | "version": "3.8.16" 389 | }, 390 | "orig_nbformat": 4 391 | }, 392 | "nbformat": 4, 393 | "nbformat_minor": 2 394 | } 395 | -------------------------------------------------------------------------------- /course_notebooks/l4-summarizing.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "attachments": {}, 5 | "cell_type": "markdown", 6 | "metadata": {}, 7 | "source": [ 8 | "[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/GitHubDaily/ChatGPT-Prompt-Engineering-for-Developers-in-Chinese/blob/master/assets/4-Summarizing/l4-summarizing.ipynb)\n", 9 | "# Summarizing\n", 10 | "In this lesson, you will summarize text with a focus on specific topics.\n", 11 | "\n", 12 | "## Setup" 13 | ] 14 | }, 15 | { 16 | "cell_type": "code", 17 | "execution_count": null, 18 | "metadata": {}, 19 | "outputs": [], 20 | "source": [ 21 | "# Basic congfig \n", 22 | "# Install basic package and set key\n", 23 | "!pip install openai\n", 24 | "!export OPENAI_API_KEY='sk-...'" 25 | ] 26 | }, 27 | { 28 | "cell_type": "code", 29 | "execution_count": null, 30 | "metadata": {}, 31 | "outputs": [], 32 | "source": [ 33 | "import openai\n", 34 | "import os\n", 35 | "\n", 36 | "from dotenv import load_dotenv, find_dotenv\n", 37 | "_ = load_dotenv(find_dotenv()) # read local .env file\n", 38 | "\n", 39 | "openai.api_key = os.getenv('OPENAI_API_KEY')" 40 | ] 41 | }, 42 | { 43 | "cell_type": "code", 44 | "execution_count": null, 45 | "metadata": {}, 46 | "outputs": [], 47 | "source": [ 48 | "def get_completion(prompt, model=\"gpt-3.5-turbo\"): # Andrew mentioned that the prompt/ completion paradigm is preferable for this class\n", 49 | " messages = [{\"role\": \"user\", \"content\": prompt}]\n", 50 | " response = openai.ChatCompletion.create(\n", 51 | " model=model,\n", 52 | " messages=messages,\n", 53 | " temperature=0, # this is the degree of randomness of the model's output\n", 54 | " )\n", 55 | " return response.choices[0].message[\"content\"]" 56 | ] 57 | }, 58 | { 59 | "attachments": {}, 60 | "cell_type": "markdown", 61 | "metadata": {}, 62 | "source": [ 63 | "## Text to summarize" 64 | ] 65 | }, 66 | { 67 | "cell_type": "code", 68 | "execution_count": null, 69 | "metadata": {}, 70 | "outputs": [], 71 | "source": [ 72 | "prod_review = \"\"\"\n", 73 | "Got this panda plush toy for my daughter's birthday, \\\n", 74 | "who loves it and takes it everywhere. It's soft and \\ \n", 75 | "super cute, and its face has a friendly look. It's \\ \n", 76 | "a bit small for what I paid though. I think there \\ \n", 77 | "might be other options that are bigger for the \\ \n", 78 | "same price. It arrived a day earlier than expected, \\ \n", 79 | "so I got to play with it myself before I gave it \\ \n", 80 | "to her.\n", 81 | "\"\"\"" 82 | ] 83 | }, 84 | { 85 | "attachments": {}, 86 | "cell_type": "markdown", 87 | "metadata": {}, 88 | "source": [ 89 | "## Summarize with a word/sentence/character limit" 90 | ] 91 | }, 92 | { 93 | "cell_type": "code", 94 | "execution_count": null, 95 | "metadata": {}, 96 | "outputs": [], 97 | "source": [ 98 | "prompt = f\"\"\"\n", 99 | "Your task is to generate a short summary of a product \\\n", 100 | "review from an ecommerce site. \n", 101 | "\n", 102 | "Summarize the review below, delimited by triple \n", 103 | "backticks, in at most 30 words. \n", 104 | "\n", 105 | "Review: ```{prod_review}```\n", 106 | "\"\"\"\n", 107 | "\n", 108 | "response = get_completion(prompt)\n", 109 | "print(response)" 110 | ] 111 | }, 112 | { 113 | "attachments": {}, 114 | "cell_type": "markdown", 115 | "metadata": {}, 116 | "source": [ 117 | "## Summarize with a focus on shipping and delivery" 118 | ] 119 | }, 120 | { 121 | "cell_type": "code", 122 | "execution_count": null, 123 | "metadata": {}, 124 | "outputs": [], 125 | "source": [ 126 | "prompt = f\"\"\"\n", 127 | "Your task is to generate a short summary of a product \\\n", 128 | "review from an ecommerce site to give feedback to the \\\n", 129 | "Shipping deparmtment. \n", 130 | "\n", 131 | "Summarize the review below, delimited by triple \n", 132 | "backticks, in at most 30 words, and focusing on any aspects \\\n", 133 | "that mention shipping and delivery of the product. \n", 134 | "\n", 135 | "Review: ```{prod_review}```\n", 136 | "\"\"\"\n", 137 | "\n", 138 | "response = get_completion(prompt)\n", 139 | "print(response)" 140 | ] 141 | }, 142 | { 143 | "attachments": {}, 144 | "cell_type": "markdown", 145 | "metadata": {}, 146 | "source": [ 147 | "## Summarize with a focus on price and value" 148 | ] 149 | }, 150 | { 151 | "cell_type": "code", 152 | "execution_count": null, 153 | "metadata": {}, 154 | "outputs": [], 155 | "source": [ 156 | "prompt = f\"\"\"\n", 157 | "Your task is to generate a short summary of a product \\\n", 158 | "review from an ecommerce site to give feedback to the \\\n", 159 | "Shipping deparmtment. \n", 160 | "\n", 161 | "Summarize the review below, delimited by triple \n", 162 | "backticks, in at most 30 words, and focusing on any aspects \\\n", 163 | "that mention shipping and delivery of the product. \n", 164 | "\n", 165 | "Review: ```{prod_review}```\n", 166 | "\"\"\"\n", 167 | "\n", 168 | "response = get_completion(prompt)\n", 169 | "print(response)" 170 | ] 171 | }, 172 | { 173 | "attachments": {}, 174 | "cell_type": "markdown", 175 | "metadata": {}, 176 | "source": [ 177 | "## Summarize with a focus on price and value" 178 | ] 179 | }, 180 | { 181 | "cell_type": "code", 182 | "execution_count": null, 183 | "metadata": {}, 184 | "outputs": [], 185 | "source": [ 186 | "prompt = f\"\"\"\n", 187 | "Your task is to generate a short summary of a product \\\n", 188 | "review from an ecommerce site to give feedback to the \\\n", 189 | "pricing deparmtment, responsible for determining the \\\n", 190 | "price of the product. \n", 191 | "\n", 192 | "Summarize the review below, delimited by triple \n", 193 | "backticks, in at most 30 words, and focusing on any aspects \\\n", 194 | "that are relevant to the price and perceived value. \n", 195 | "\n", 196 | "Review: ```{prod_review}```\n", 197 | "\"\"\"\n", 198 | "\n", 199 | "response = get_completion(prompt)\n", 200 | "print(response)" 201 | ] 202 | }, 203 | { 204 | "attachments": {}, 205 | "cell_type": "markdown", 206 | "metadata": {}, 207 | "source": [ 208 | "#### Comment\n", 209 | "- Summaries include topics that are not related to the topic of focus." 210 | ] 211 | }, 212 | { 213 | "attachments": {}, 214 | "cell_type": "markdown", 215 | "metadata": {}, 216 | "source": [ 217 | "## Try \"extract\" instead of \"summarize\"" 218 | ] 219 | }, 220 | { 221 | "cell_type": "code", 222 | "execution_count": null, 223 | "metadata": {}, 224 | "outputs": [], 225 | "source": [ 226 | "prompt = f\"\"\"\n", 227 | "Your task is to extract relevant information from \\ \n", 228 | "a product review from an ecommerce site to give \\\n", 229 | "feedback to the Shipping department. \n", 230 | "\n", 231 | "From the review below, delimited by triple quotes \\\n", 232 | "extract the information relevant to shipping and \\ \n", 233 | "delivery. Limit to 30 words. \n", 234 | "\n", 235 | "Review: ```{prod_review}```\n", 236 | "\"\"\"\n", 237 | "\n", 238 | "response = get_completion(prompt)\n", 239 | "print(response)" 240 | ] 241 | }, 242 | { 243 | "attachments": {}, 244 | "cell_type": "markdown", 245 | "metadata": {}, 246 | "source": [ 247 | "## Summarize multiple product reviews" 248 | ] 249 | }, 250 | { 251 | "cell_type": "code", 252 | "execution_count": null, 253 | "metadata": {}, 254 | "outputs": [], 255 | "source": [ 256 | "review_1 = prod_review \n", 257 | "\n", 258 | "# review for a standing lamp\n", 259 | "review_2 = \"\"\"\n", 260 | "Needed a nice lamp for my bedroom, and this one \\\n", 261 | "had additional storage and not too high of a price \\\n", 262 | "point. Got it fast - arrived in 2 days. The string \\\n", 263 | "to the lamp broke during the transit and the company \\\n", 264 | "happily sent over a new one. Came within a few days \\\n", 265 | "as well. It was easy to put together. Then I had a \\\n", 266 | "missing part, so I contacted their support and they \\\n", 267 | "very quickly got me the missing piece! Seems to me \\\n", 268 | "to be a great company that cares about their customers \\\n", 269 | "and products. \n", 270 | "\"\"\"\n", 271 | "\n", 272 | "# review for an electric toothbrush\n", 273 | "review_3 = \"\"\"\n", 274 | "My dental hygienist recommended an electric toothbrush, \\\n", 275 | "which is why I got this. The battery life seems to be \\\n", 276 | "pretty impressive so far. After initial charging and \\\n", 277 | "leaving the charger plugged in for the first week to \\\n", 278 | "condition the battery, I've unplugged the charger and \\\n", 279 | "been using it for twice daily brushing for the last \\\n", 280 | "3 weeks all on the same charge. But the toothbrush head \\\n", 281 | "is too small. I’ve seen baby toothbrushes bigger than \\\n", 282 | "this one. I wish the head was bigger with different \\\n", 283 | "length bristles to get between teeth better because \\\n", 284 | "this one doesn’t. Overall if you can get this one \\\n", 285 | "around the $50 mark, it's a good deal. The manufactuer's \\\n", 286 | "replacements heads are pretty expensive, but you can \\\n", 287 | "get generic ones that're more reasonably priced. This \\\n", 288 | "toothbrush makes me feel like I've been to the dentist \\\n", 289 | "every day. My teeth feel sparkly clean! \n", 290 | "\"\"\"\n", 291 | "\n", 292 | "# review for a blender\n", 293 | "review_4 = \"\"\"\n", 294 | "So, they still had the 17 piece system on seasonal \\\n", 295 | "sale for around $49 in the month of November, about \\\n", 296 | "half off, but for some reason (call it price gouging) \\\n", 297 | "around the second week of December the prices all went \\\n", 298 | "up to about anywhere from between $70-$89 for the same \\\n", 299 | "system. And the 11 piece system went up around $10 or \\\n", 300 | "so in price also from the earlier sale price of $29. \\\n", 301 | "So it looks okay, but if you look at the base, the part \\\n", 302 | "where the blade locks into place doesn’t look as good \\\n", 303 | "as in previous editions from a few years ago, but I \\\n", 304 | "plan to be very gentle with it (example, I crush \\\n", 305 | "very hard items like beans, ice, rice, etc. in the \\ \n", 306 | "blender first then pulverize them in the serving size \\\n", 307 | "I want in the blender then switch to the whipping \\\n", 308 | "blade for a finer flour, and use the cross cutting blade \\\n", 309 | "first when making smoothies, then use the flat blade \\\n", 310 | "if I need them finer/less pulpy). Special tip when making \\\n", 311 | "smoothies, finely cut and freeze the fruits and \\\n", 312 | "vegetables (if using spinach-lightly stew soften the \\ \n", 313 | "spinach then freeze until ready for use-and if making \\\n", 314 | "sorbet, use a small to medium sized food processor) \\ \n", 315 | "that you plan to use that way you can avoid adding so \\\n", 316 | "much ice if at all-when making your smoothie. \\\n", 317 | "After about a year, the motor was making a funny noise. \\\n", 318 | "I called customer service but the warranty expired \\\n", 319 | "already, so I had to buy another one. FYI: The overall \\\n", 320 | "quality has gone done in these types of products, so \\\n", 321 | "they are kind of counting on brand recognition and \\\n", 322 | "consumer loyalty to maintain sales. Got it in about \\\n", 323 | "two days.\n", 324 | "\"\"\"\n", 325 | "\n", 326 | "reviews = [review_1, review_2, review_3, review_4]\n" 327 | ] 328 | }, 329 | { 330 | "cell_type": "code", 331 | "execution_count": null, 332 | "metadata": {}, 333 | "outputs": [], 334 | "source": [ 335 | "for i in range(len(reviews)):\n", 336 | " prompt = f\"\"\"\n", 337 | " Your task is to generate a short summary of a product \\ \n", 338 | " review from an ecommerce site. \n", 339 | "\n", 340 | " Summarize the review below, delimited by triple \\\n", 341 | " backticks in at most 20 words. \n", 342 | "\n", 343 | " Review: ```{reviews[i]}```\n", 344 | " \"\"\"\n", 345 | "\n", 346 | " response = get_completion(prompt)\n", 347 | " print(i, response, \"\\n\")" 348 | ] 349 | }, 350 | { 351 | "cell_type": "code", 352 | "execution_count": null, 353 | "metadata": {}, 354 | "outputs": [], 355 | "source": [] 356 | } 357 | ], 358 | "metadata": { 359 | "kernelspec": { 360 | "display_name": "gpt_index", 361 | "language": "python", 362 | "name": "python3" 363 | }, 364 | "language_info": { 365 | "codemirror_mode": { 366 | "name": "ipython", 367 | "version": 3 368 | }, 369 | "file_extension": ".py", 370 | "mimetype": "text/x-python", 371 | "name": "python", 372 | "nbconvert_exporter": "python", 373 | "pygments_lexer": "ipython3", 374 | "version": "3.8.16" 375 | }, 376 | "orig_nbformat": 4 377 | }, 378 | "nbformat": 4, 379 | "nbformat_minor": 2 380 | } 381 | -------------------------------------------------------------------------------- /subtitles/raw/prompt_eng_04_summarizing.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:09,680 3 | 当今世界有如此多的文字，但我们几乎没有人有足够的时间 4 | There's so much text in today's world, pretty much none of us have enough time to read all 5 | 6 | 2 7 | 00:00:09,680 --> 00:00:12,760 8 | 来阅读这些文字 9 | the things we wish we had time to 10 | 11 | 3 12 | 00:00:12,760 --> 00:00:16,920 13 | 因此，我认为，大语言模型最令人兴奋的应用之一 14 | So one of the most exciting applications I've seen of large language models is to use it 15 | 16 | 4 17 | 00:00:16,920 --> 00:00:18,660 18 | 是用它来对文本进行总结 19 | to summarize text 20 | 21 | 5 22 | 00:00:18,660 --> 00:00:24,480 23 | 这也是我看到多个团队在不同软件应用中构建的东西 24 | And this is something that I'm seeing multiple teams build into multiple software applications 25 | 26 | 6 27 | 00:00:24,480 --> 00:00:26,880 28 | 你可以在ChatGPT网站上这样做 29 | You can do this in the ChatGPT web interface 30 | 31 | 7 32 | 00:00:26,880 --> 00:00:31,000 33 | 我经常用这种方式来总结文章 34 | I do this all the time to summarize articles so I can just kind of read the content of 35 | 36 | 8 37 | 00:00:31,000 --> 00:00:33,000 38 | 这样我就可以比以前多读很多文章 39 | many more articles than I previously could 40 | 41 | 9 42 | 00:00:33,000 --> 00:00:37,880 43 | 你将在本课程中，学习如何以编程方式来实现这个任务 44 | And if you want to do this more programmatically, you'll see how to in this lesson 45 | 46 | 10 47 | 00:00:37,880 --> 00:00:43,520 48 | 因此，让我们深入研究代码，看看你如何使用它来总结文本 49 | So that's let's dig into the code to see how you could use this yourself to summarize text 50 | 51 | 11 52 | 00:00:43,520 --> 00:00:48,280 53 | 让我们从之前的初始代码开始，导入openai，加载 54 | So let's start off with the same starter code as you saw before of import openai, load the 55 | 56 | 12 57 | 00:00:48,280 --> 00:00:52,620 58 | API Key，这里是get_completion辅助函数 59 | API key and here's that get_completion helper function 60 | 61 | 13 62 | 00:00:52,620 --> 00:01:00,640 63 | 我将以“总结此产品评价”的任务作为运行示例 64 | I'm going to use as the running example, the task of summarizing this product review 65 | 66 | 14 67 | 00:01:00,640 --> 00:01:04,640 68 | 我为女儿的生日买了这只熊猫毛绒玩具，女儿非常喜欢 69 | Got this panda plush toy for my daughter's birthday 70 | 71 | 15 72 | 00:01:04,640 --> 00:01:06,320 73 | 无论去哪儿都要带上它，诸如此类 74 | and so on and so on 75 | 76 | 16 77 | 00:01:06,320 --> 00:01:13,640 78 | 如果你正在建立一个电子商务网站，并且有大量的评论 79 | If you're building an e commerce website, and there's just a large volume of reviews 80 | 81 | 17 82 | 00:01:13,640 --> 00:01:18,680 83 | 有一个工具可以用来总结冗长的评论，给你提供一种方式 84 | having a tool to summarize the lengthy reviews could give you a way to very quickly glance 85 | 86 | 18 87 | 00:01:18,680 --> 00:01:24,200 88 | 让你可以浏览更多评论，更好地了解所有客户的想法 89 | over more reviews to get a better sense of what all your customers are thinking 90 | 91 | 19 92 | 00:01:24,200 --> 00:01:27,280 93 | 因此，这里有一个生成摘要的提示 94 | So here's a prompt for generating a summary 95 | 96 | 20 97 | 00:01:27,280 --> 00:01:30,600 98 | 你的任务是对电子商务网站上的产品评论生成一个简短的总结，总结如下 99 | Your task is generate a short summary of a product review from e commerce website, summarize 100 | 101 | 21 102 | 00:01:30,600 --> 00:01:36,800 103 | 下面的评论等等，最多只能用30个字 104 | review below, and so on, in at most 30 words 105 | 106 | 22 107 | 00:01:36,800 --> 00:01:42,400 108 | 因此，这个柔软可爱的熊猫毛绒玩具受到女儿的喜爱，但价格有点小贵 109 | And so this is soft and cute panda plush toy loved by daughter, but small for the price 110 | 111 | 23 112 | 00:01:42,400 --> 00:01:43,400 113 | 提前到货 114 | arrived early 115 | 116 | 24 117 | 00:01:43,400 --> 00:01:44,400 118 | 还不错 119 | Not bad 120 | 121 | 25 122 | 00:01:44,400 --> 00:01:45,400 123 | 这是个相当好的总结 124 | It's a pretty good summary 125 | 126 | 26 127 | 00:01:45,400 --> 00:01:49,760 128 | 正如你在前面的视频中看到的，你还可以玩一些东西，比如控制 129 | And as you saw in the previous video, you can also play with things like controlling 130 | 131 | 27 132 | 00:01:49,760 --> 00:01:55,720 133 | 字符数或句子数来影响这个摘要的长度 134 | the character count or the number of sentences to affect the length of this summary 135 | 136 | 28 137 | 00:01:55,720 --> 00:01:59,920 138 | 现在，有时在创建一个摘要时，如果你有一个非常具体的目的 139 | Now sometimes when creating a summary, if you have a very specific purpose in mind for 140 | 141 | 29 142 | 00:01:59,920 --> 00:02:05,600 143 | 例如，你想给运输部门提供反馈 144 | the summary, for example, if you want to give feedback to the shipping department, you can 145 | 146 | 30 147 | 00:02:05,600 --> 00:02:11,200 148 | 你也可以修改提示来突出这点，这样他们就可以生成一个更好的总结 149 | also modify the prompt to reflect that so that they can generate a summary that is more 150 | 151 | 31 152 | 00:02:11,200 --> 00:02:15,600 153 | 适用于你业务中的一个特定群体 154 | applicable to one particular group in your business 155 | 156 | 32 157 | 00:02:15,600 --> 00:02:26,880 158 | 例如，如果我想要给运输部门反馈 159 | So for example, if I add to give feedback to the shipping department, let's say I change 160 | 161 | 33 162 | 00:02:26,880 --> 00:02:35,800 163 | 那么我会将关注点集中在商品的运输和交付等方面 164 | this to start to focus on any aspects that mention shipping and delivery of the product 165 | 166 | 34 167 | 00:02:35,800 --> 00:02:42,560 168 | 如果我运行这个指令，那么你会得到一个新的摘要 169 | And if I run this, then again, you get a summary, but instead of starting off with soft and 170 | 171 | 35 172 | 00:02:42,560 --> 00:02:49,120 173 | 这次不是以“柔软可爱的熊猫毛绒玩具”开始，而是强调它比预期提前了一天送达 174 | cute panda plush toy, it now focuses on the fact that it arrived a day earlier than expected 175 | 176 | 36 177 | 00:02:49,120 --> 00:02:53,600 178 | 然后它仍然有其他细节 179 | And then it still has, you know, other details 180 | 181 | 37 182 | 00:02:53,600 --> 00:02:59,640 183 | 或者作为另一个例子，如果我们不是想给运输部门提供反馈 184 | Or as another example, if we aren't trying to give feedback to the shipping department 185 | 186 | 38 187 | 00:02:59,640 --> 00:03:09,520 188 | 而是给定价部门提供反馈 189 | but let's say we want to give feedback to the pricing department 190 | 191 | 39 192 | 00:03:09,520 --> 00:03:15,760 193 | 定价部门负责确定产品的价格 194 | So the pricing department is responsible for determining the price of the product 195 | 196 | 40 197 | 00:03:15,760 --> 00:03:22,960 198 | 我会告诉它专注于与产品价格和感知价值相关的任何方面 199 | And I'm going to tell it to focus on any aspects that are relevant to the price and perceived 200 | 201 | 41 202 | 00:03:22,960 --> 00:03:25,240 203 | 价值 204 | value 205 | 206 | 42 207 | 00:03:25,240 --> 00:03:31,520 208 | 那么这将生成一个不同的总结，它说 209 | Then this generates a different summary that it says 210 | 211 | 43 212 | 00:03:31,520 --> 00:03:33,480 213 | 也许这个尺寸的价格可能太高了 214 | maybe the price may be too high for a size 215 | 216 | 44 217 | 00:03:33,480 --> 00:03:38,640 218 | 现在，在我为运输部门或定价部门生成的摘要中 219 | Now in the summaries that I've generated for the shipping department or the pricing department 220 | 221 | 45 222 | 00:03:38,640 --> 00:03:44,400 223 | 它更侧重于与这些特定部门有关的信息 224 | it focuses a bit more on information relevant to those specific departments 225 | 226 | 46 227 | 00:03:44,400 --> 00:03:50,000 228 | 事实上，你可以随时暂停视频 229 | And in fact, feel free to pause the video now and maybe ask it to generate information 230 | 231 | 47 232 | 00:03:50,000 --> 00:03:55,360 233 | 然后让它为负责产品客户体验的部门提供有关产品的信息 234 | for the product department responsible for the customer experience of the product or 235 | 236 | 48 237 | 00:03:55,360 --> 00:04:01,920 238 | 或者为你认为对电子商务站点有趣的其他某些方面提供信息 239 | for something else that you think might be interesting to an e-commerce site 240 | 241 | 49 242 | 00:04:01,920 --> 00:04:07,000 243 | 但在这些摘要中，尽管它生成了与运输有关的信息 244 | But in these summaries, even though it generated the information relevant to shipping, it had 245 | 246 | 50 247 | 00:04:07,000 --> 00:04:12,240 248 | 但也有一些其他信息，你可以决定这些信息是否有帮助 249 | some other information too, which you could decide may or may not be helpful 250 | 251 | 51 252 | 00:04:12,240 --> 00:04:20,120 253 | 因此，根据你想要总结的方式，你也可以要求它提取信息 254 | So depending on how you want to summarize it, you can also ask it to extract information 255 | 256 | 52 257 | 00:04:20,120 --> 00:04:22,960 258 | 而不是对其进行总结 259 | rather than summarize it 260 | 261 | 53 262 | 00:04:22,960 --> 00:04:27,920 263 | 因此，这里有一个提示，说你的任务是 264 | So here's a prompt that says you're tasked to extract relevant information to give feedback 265 | 266 | 54 267 | 00:04:27,920 --> 00:04:29,200 268 | 提取相关的信息来反馈给给运输部门 269 | to the shipping department 270 | 271 | 55 272 | 00:04:29,200 --> 00:04:35,440 273 | 现在它只是说，产品比预期早到了一天，没有其他的信息 274 | And now it just says, product arrived a day earlier than expected without all of the other 275 | 276 | 56 277 | 00:04:35,440 --> 00:04:41,240 278 | 这在一般的总结中也是有帮助的，但如果想知道运输方面发生了什么 279 | information, which was also helpful in a general summary, but less specific to the shipping 280 | 281 | 57 282 | 00:04:41,240 --> 00:04:45,000 283 | 就不那么具体了 284 | department if all it wants to know is what happened with the shipping 285 | 286 | 58 287 | 00:04:45,000 --> 00:04:50,600 288 | 最后，让我与你分享一个具体的例子，说明如何在工作流程中使用这个方法 289 | Lastly, let me just share with you a concrete example for how to use this in a workflow 290 | 291 | 59 292 | 00:04:50,600 --> 00:04:56,760 293 | 来帮助总结多个评论，使其更容易阅读 294 | to help summarize multiple reviews to make them easier to read 295 | 296 | 60 297 | 00:04:56,760 --> 00:04:59,920 298 | 因此，这里有几条评论 299 | So here are a few reviews 300 | 301 | 61 302 | 00:04:59,920 --> 00:05:03,920 303 | 这有点长，但这是对落地灯的第二个评论 304 | This is kind of long, but here's a second review for a standing lamp 305 | 306 | 62 307 | 00:05:03,920 --> 00:05:04,920 308 | 需要一盏灯放在卧室里 309 | need a lamp on a bedroom 310 | 311 | 63 312 | 00:05:04,920 --> 00:05:07,280 313 | 这是关于电动牙刷的第三条评论 314 | Here's a third review for an electric toothbrush 315 | 316 | 64 317 | 00:05:07,280 --> 00:05:09,600 318 | 我的牙科保健师推荐的 319 | My dental hygienist recommended 320 | 321 | 65 322 | 00:05:09,600 --> 00:05:12,640 323 | 关于电动牙刷的评论有点长 324 | Kind of a long review about the electric toothbrush 325 | 326 | 66 327 | 00:05:12,640 --> 00:05:17,920 328 | 这是关于搅拌机的评论，当时它说是季节性销售的17件套装系统 329 | This is a review for a blender when it said, so said 17 piece system on seasonal sale and 330 | 331 | 67 332 | 00:05:17,920 --> 00:05:18,920 333 | 诸如此类 334 | so on and so on 335 | 336 | 68 337 | 00:05:18,920 --> 00:05:20,280 338 | 这实际上是很多的文字 339 | This is actually a lot of text 340 | 341 | 69 342 | 00:05:20,280 --> 00:05:23,280 343 | 如果你愿意，可以随意暂停视频，读完这些文字 344 | If you want, feel free to pause the video and read through all this text 345 | 346 | 70 347 | 00:05:23,280 --> 00:05:28,480 348 | 但是，如果你想知道这些评论者写了什么，而不必停下来阅读所有的 349 | But what if you want to know what these reviewers wrote without having to stop and read all 350 | 351 | 71 352 | 00:05:28,480 --> 00:05:29,480 353 | 这方面的细节呢？ 354 | this in detail? 355 | 356 | 72 357 | 00:05:29,480 --> 00:05:41,600 358 | 所以我要把 review_1 设为我们之前展示的那个产品评论 359 | So I'm going to set review_1 to be just the product review that we had up there 360 | 361 | 73 362 | 00:05:41,600 --> 00:05:45,320 363 | 我将把所有这些评论放到一个列表中 364 | I'm going to put all of these reviews into a list 365 | 366 | 74 367 | 00:05:45,320 --> 00:05:53,800 368 | 现在，如果我对这些评论实施一个for循环 369 | And now if I implement a for loop over the reviews 370 | 371 | 75 372 | 00:05:53,800 --> 00:06:01,240 373 | 这是我的提示，我要求它最多用20个字来总结它 374 | So here's my prompt and here I've asked it to summarize it in at most 20 words 375 | 376 | 76 377 | 00:06:01,240 --> 00:06:07,320 378 | 然后让它得到响应并打印出来 379 | Then let's have it get the response and print it out 380 | 381 | 77 382 | 00:06:07,320 --> 00:06:10,760 383 | 让我们运行这个程序 384 | And let's run that 385 | 386 | 78 387 | 00:06:10,760 --> 00:06:16,360 388 | 它打印出的第一条评论是熊猫玩具的评论，灯的总结评论 389 | And it prints out the first review is that PandaToy review, summary review of the lamp 390 | 391 | 79 392 | 00:06:16,360 --> 00:06:21,880 393 | 牙刷的摘要评论，然后是搅拌器 394 | summary review of the toothbrush, and then the blender 395 | 396 | 80 397 | 00:06:21,880 --> 00:06:29,800 398 | 因此，如果你有一个网站，你有数百条评论，你可以想象你如何 399 | And so if you have a website where you have hundreds of reviews, you can imagine how you 400 | 401 | 81 402 | 00:06:29,800 --> 00:06:35,360 403 | 使用它来建立一个控制面板，为大量的评论，生成简短的摘要 404 | might use this to build a dashboard, to take huge numbers of reviews, generate short summaries 405 | 406 | 82 407 | 00:06:35,360 --> 00:06:40,760 408 | 这样你或其他人就可以更快地浏览这些评论 409 | of them so that you or someone else can browse the reviews much more quickly 410 | 411 | 83 412 | 00:06:40,760 --> 00:06:46,200 413 | 然后，如果他们愿意，可以点击查看原始的长篇评论 414 | And then if they wish, maybe click in to see the original longer review 415 | 416 | 84 417 | 00:06:46,200 --> 00:06:51,640 418 | 这可以帮助你更高效地了解所有客户的想法 419 | And this can help you efficiently get a better sense of what all of your customers are thinking 420 | 421 | 85 422 | 00:06:51,640 --> 00:06:55,640 423 | 好了，总结就到这里 424 | All right, so that's it for summarizing 425 | 426 | 86 427 | 00:06:55,640 --> 00:07:00,960 428 | 我希望你能想象一下，如果你有任何有许多文本的应用 429 | And I hope that you can picture if you have any applications with many pieces of text 430 | 431 | 87 432 | 00:07:00,960 --> 00:07:05,680 433 | 你可以使用这样的提示来总结，帮助人们快速了解文字中的内容和各种文本 434 | how you can use prompts like these to summarize them to help people quickly get a sense of 435 | 436 | 88 437 | 00:07:05,680 --> 00:07:10,240 438 | 可以选择进一步深入挖掘 439 | what's in the text, the many pieces of text, and perhaps optionally dig in more 440 | 441 | 89 442 | 00:07:10,240 --> 00:07:12,520 443 | 如果有需要的话 444 | if they wish 445 | 446 | 90 447 | 00:07:12,520 --> 00:07:17,640 448 | 在下一个视频中，我们将看到大语言模型的另一个能力，即 449 | In the next video, we'll look at another capability of large language models, which is to make 450 | 451 | 91 452 | 00:07:17,640 --> 00:07:19,880 453 | 推断的能力 454 | inferences using text 455 | 456 | 92 457 | 00:07:19,880 --> 00:07:24,280 458 | 例如，如果你有一些产品评论数据 459 | For example, what if you had, again, product reviews, and you wanted to very quickly get 460 | 461 | 93 462 | 00:07:24,280 --> 00:07:29,680 463 | 你希望快速了解哪些评论具有积极或消极情感，该怎么办？ 464 | a sense of which product reviews have a positive or a negative sentiment? 465 | 466 | 94 467 | 00:07:29,680 --> 00:07:49,840 468 | 让我们在下一个视频中看看如何实现这个功能 469 | Let's take a look at how to do that in the next video 470 | 471 | -------------------------------------------------------------------------------- /subtitles/raw/prompt_eng_01_intro.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:09,000 3 | 欢迎学习《面向开发者的 ChatGPT Prompt 工程》课程 4 | Welcome to this course on ChatGPT Prompt Engineering for Developers 5 | 6 | 2 7 | 00:00:09,000 --> 00:00:14,000 8 | 很高兴能邀请到 Isa Fulford 和我一起授课 9 | I'm thrilled to have with me Iza Fulford to teach this along with me 10 | 11 | 3 12 | 00:00:14,000 --> 00:00:18,000 13 | 她是 OpenAI 的技术人员之一，曾建立了流行的 14 | She is a member of the technical staff of OpenAI and had built the popular 15 | 16 | 4 17 | 00:00:18,000 --> 00:00:23,000 18 | ChatGPT 检索插件，她大部分工作是教人们 19 | ChatGPT Retrieval plugin and a large part of her work has been teaching people 20 | 21 | 5 22 | 00:00:23,000 --> 00:00:27,000 23 | 如何在产品中使用LLM或大语言模型技术 24 | how to use LLM or large language model technology in products 25 | 26 | 6 27 | 00:00:27,000 --> 00:00:31,000 28 | 她还为OpenAI手册做出了贡献，教会人们如何使用提示 29 | She's also contributed to the OpenAI cookbook that teaches people prompting 30 | 31 | 7 32 | 00:00:31,000 --> 00:00:32,000 33 | 所以，很高兴有你参与 34 | So thrilled to have you with you 35 | 36 | 8 37 | 00:00:32,000 --> 00:00:37,000 38 | 我也很高兴能在这里与大家分享一些提示的最佳实践 39 | And I'm thrilled to be here and share some prompting best practices with you all 40 | 41 | 9 42 | 00:00:37,000 --> 00:00:42,000 43 | 在互联网上有很多关于提示资料以及一些文章 44 | So there's been a lot of material on the internet for prompting with articles 45 | 46 | 10 47 | 00:00:42,000 --> 00:00:45,000 48 | 比如，每个人都必须知道的30个提示 49 | like 30 prompts everyone has to know 50 | 51 | 11 52 | 00:00:45,000 --> 00:00:50,000 53 | 许多人正在使用 ChatGPT 的 Web 用户界面 54 | A lot of that has been focused on the ChatGPT web user interface 55 | 56 | 12 57 | 00:00:50,000 --> 00:00:54,000 58 | 来完成特定而且通常是一次性的任务 59 | which many people are using to do specific and often one-off tasks 60 | 61 | 13 62 | 00:00:54,000 --> 00:01:00,000 63 | 但作为一名开发者，我认为LLMs和大语言模型的强大也是不可忽视的 64 | But I think the power of LLMs, large language models as a developer too 65 | 66 | 14 67 | 00:01:00,000 --> 00:01:04,000 68 | 那就是通过API调用LLM来快速构建软件应用程序 69 | that is using API calls to LLMs to quickly build software applications 70 | 71 | 15 72 | 00:01:04,000 --> 00:01:08,000 73 | 我认为这仍然被严重低估 74 | I think that is still very underappreciated 75 | 76 | 16 77 | 00:01:08,000 --> 00:01:12,000 78 | 事实上，我在AI Fund的团队，也就是Deep Learning.AI的姐妹公司 79 | In fact, my team at AI Fund, which is a sister company to Deep Learning.AI 80 | 81 | 17 82 | 00:01:12,000 --> 00:01:16,000 83 | 一直在与许多初创公司、不同项目合作 84 | has been working with many startups on applying these technologies 85 | 86 | 18 87 | 00:01:16,000 --> 00:01:18,000 88 | 来应用这些技术 89 | to many different applications 90 | 91 | 19 92 | 00:01:18,000 --> 00:01:23,000 93 | 看到LLM API能够让开发者快速建立一些东西 94 | And it's been exciting to see what LLM APIs can enable developers 95 | 96 | 20 97 | 00:01:23,000 --> 00:01:25,000 98 | 这很令人兴奋 99 | to very quickly build 100 | 101 | 21 102 | 00:01:25,000 --> 00:01:29,000 103 | 所以在这个课程中，我们将与你分享一些可能性 104 | So in this course, we'll share with you some of the possibilities 105 | 106 | 22 107 | 00:01:29,000 --> 00:01:34,000 108 | 你可以做什么，以及如何做的最佳实践 109 | for what you can do as well as best practices for how you can do them 110 | 111 | 23 112 | 00:01:34,000 --> 00:01:36,000 113 | 这里会覆盖很多资料 114 | There's a lot of material to cover 115 | 116 | 24 117 | 00:01:36,000 --> 00:01:41,000 118 | 首先，你将学习一些软件开发的提示词最佳实践 119 | First, you'll learn some prompting best practices for software development 120 | 121 | 25 122 | 00:01:41,000 --> 00:01:45,000 123 | 然后，我们将涵盖一些常见的用例、总结、推断 124 | Then we'll cover some common use cases, summarizing, inferring 125 | 126 | 26 127 | 00:01:45,000 --> 00:01:50,000 128 | 转化、扩展，然后使用LLM建立一个聊天机器人 129 | transforming, expanding, and then you'll build a chatbot using an LLM 130 | 131 | 27 132 | 00:01:50,000 --> 00:01:53,000 133 | 我们希望这将激发你 134 | We hope that this will spark your imagination about new applications 135 | 136 | 28 137 | 00:01:53,000 --> 00:01:55,000 138 | 开发新应用的想象力 139 | that you can build 140 | 141 | 29 142 | 00:01:55,000 --> 00:01:58,000 143 | 因此，在大型语言模型或LLM的发展中 144 | So in the development of large language models or LLMs 145 | 146 | 30 147 | 00:01:58,000 --> 00:02:02,000 148 | 大体上有两种类型的LLM，我把它们称为 149 | there have been broadly two types of LLMs, which I'm going to refer to 150 | 151 | 31 152 | 00:02:02,000 --> 00:02:06,000 153 | 基础LLM和指令学习LLM 154 | as base LLMs and instruction-tuned LLMs 155 | 156 | 32 157 | 00:02:06,000 --> 00:02:11,000 158 | 基础LLM已经被训练成基于文本训练数据来预测下一个单词 159 | So base LLM has been trained to predict the next word based on text training data 160 | 161 | 33 162 | 00:02:11,000 --> 00:02:15,000 163 | 通常通过互联网和其他来源训练大量数据 164 | often trained on a large amount of data from the internet and other sources 165 | 166 | 34 167 | 00:02:15,000 --> 00:02:19,000 168 | 并计算出下一个最可能出现的词是什么 169 | to figure out what's the next most likely word to follow 170 | 171 | 35 172 | 00:02:19,000 --> 00:02:24,000 173 | 比如，你输入这个提示："从前有一只独角兽" 174 | So for example, if you were to prompt this, once upon a time there was a unicorn 175 | 176 | 36 177 | 00:02:24,000 --> 00:02:28,000 178 | 它会进行补全，并预测接下来的几个词是 179 | it may complete this, that is, it may predict the next several words are 180 | 181 | 37 182 | 00:02:28,000 --> 00:02:31,000 183 | "和所有独角兽朋友一起生活在一个神奇的森林里" 184 | that live in a magical forest with all unicorn friends 185 | 186 | 38 187 | 00:02:31,000 --> 00:02:35,000 188 | 但是如果你是用"法国的首都是什么"作为提示 189 | But if you were to prompt this with what is the capital of France 190 | 191 | 39 192 | 00:02:35,000 --> 00:02:40,000 193 | 那么根据互联网上的文章 194 | then based on what articles on the internet might have 195 | 196 | 40 197 | 00:02:40,000 --> 00:02:44,000 198 | 很有可能是，基础LLM会以下列方式完成这个任务 199 | it's quite possible that the base LLM will complete this with 200 | 201 | 41 202 | 00:02:44,000 --> 00:02:48,000 203 | 什么是法国最大的城市，什么是法国的人口，等等 204 | what is France's largest city, what is France's population, and so on 205 | 206 | 42 207 | 00:02:48,000 --> 00:02:52,000 208 | 因为互联网上的文章，可能会列出 209 | because articles on the internet could quite plausibly be lists of quiz questions 210 | 211 | 43 212 | 00:02:52,000 --> 00:02:55,000 213 | 关于法国的小测验问题列表 214 | about the country of France 215 | 216 | 44 217 | 00:02:55,000 --> 00:03:00,000 218 | 与此相反，指令学习LLM，发展势头较猛 219 | In contrast, an instruction-tuned LLM, which is where a lot of momentum 220 | 221 | 45 222 | 00:03:00,000 --> 00:03:04,000 223 | LLM的研究和实践一直在进行 224 | of LLM research and practice has been going 225 | 226 | 46 227 | 00:03:04,000 --> 00:03:08,000 228 | 一个经过指令学习的LLM已经被训练得能够遵循指令 229 | an instruction-tuned LLM has been trained to follow instructions 230 | 231 | 47 232 | 00:03:08,000 --> 00:03:11,000 233 | 因此，如果你问它，法国的首都是什么？ 234 | So if you were to ask it, what is the capital of France 235 | 236 | 48 237 | 00:03:11,000 --> 00:03:15,000 238 | 它很可能输出法国的首都是巴黎 239 | it's much more likely to output something like the capital of France is Paris 240 | 241 | 49 242 | 00:03:15,000 --> 00:03:19,000 243 | 因此，指令学习的LLM的典型训练方式是 244 | So the way that instruction-tuned LLMs are typically trained is 245 | 246 | 50 247 | 00:03:19,000 --> 00:03:23,000 248 | 从一个在大量文本数据上训练过的基础LLM开始 249 | you start off with a base LLM that's been trained on a huge amount of text data 250 | 251 | 51 252 | 00:03:23,000 --> 00:03:28,000 253 | 然后进一步训练它，用输入和输出来进一步微调它 254 | and further train it, further fine-tune it with inputs and outputs 255 | 256 | 52 257 | 00:03:28,000 --> 00:03:32,000 258 | 这些输入和输出都是指令，也是遵循这些指令的良好尝试 259 | that are instructions and good attempts to follow those instructions 260 | 261 | 53 262 | 00:03:32,000 --> 00:03:36,000 263 | 然后经常使用一种叫做RLHF的技术进一步完善 264 | and then often further refine using a technique called RLHF 265 | 266 | 54 267 | 00:03:36,000 --> 00:03:41,000 268 | 从人类反馈中进行强化学习，以使系统能够更好地 269 | reinforcement learning from human feedback, to make the system better able 270 | 271 | 55 272 | 00:03:41,000 --> 00:03:43,000 273 | 提供帮助并遵循指令 274 | to be helpful and follow instructions 275 | 276 | 56 277 | 00:03:43,000 --> 00:03:47,000 278 | 因为经过指令学习的LLM已经被训练得很有帮助 279 | Because instruction-tuned LLMs have been trained to be helpful 280 | 281 | 57 282 | 00:03:47,000 --> 00:03:51,000 283 | 诚实且无害，因此，举例来说，它们不太可能输出 284 | honest, and harmless, so for example, they're less likely to output 285 | 286 | 58 287 | 00:03:51,000 --> 00:03:55,000 288 | 那些与基础LLM相比，会出问题的文本，如有害的输出 289 | problematic text, such as toxic outputs, compared to base LLM 290 | 291 | 59 292 | 00:03:55,000 --> 00:03:59,000 293 | 很多实际应用场景已经开始向 294 | a lot of the practical usage scenarios have been shifting 295 | 296 | 60 297 | 00:03:59,000 --> 00:04:01,000 298 | 指令学习LLM转移 299 | toward instruction-tuned LLMs 300 | 301 | 61 302 | 00:04:01,000 --> 00:04:04,000 303 | 你在互联网上找到的一些最佳实践可能更适用于基础LLM 304 | Some of the best practices you find on the internet may be more suited 305 | 306 | 62 307 | 00:04:04,000 --> 00:04:08,000 308 | 但对于今天的大多数实际应用情况来说，它们可能不太合适 309 | for a base LLM, but for most practical applications today 310 | 311 | 63 312 | 00:04:08,000 --> 00:04:13,000 313 | 我们建议大多数人多关注指令学习LLM 314 | we would recommend most people instead focus on instruction-tuned LLMs 315 | 316 | 64 317 | 00:04:13,000 --> 00:04:17,000 318 | 它们更容易使用，而且由于OpenAI 319 | which are easier to use, and also because of the work of OpenAI 320 | 321 | 65 322 | 00:04:17,000 --> 00:04:22,000 323 | 和其他LLM公司的工作，也将变得更加安全和一致 324 | and other LLM companies, becoming safer and more aligned 325 | 326 | 66 327 | 00:04:22,000 --> 00:04:27,000 328 | 因此，本课程将重点介绍指令学习LLM的最佳实践 329 | So this course will focus on best practices for instruction-tuned LLMs 330 | 331 | 67 332 | 00:04:27,000 --> 00:04:32,000 333 | 这也是我们建议你在大多数应用中使用的 334 | which is what we recommend you use for most of your applications 335 | 336 | 68 337 | 00:04:32,000 --> 00:04:36,000 338 | 在继续之前，我想感谢来自OpenAI 339 | Before moving on, I just want to acknowledge the team from OpenAI 340 | 341 | 69 342 | 00:04:36,000 --> 00:04:39,000 343 | 和DeepLearning.ai的团队 344 | and DeepLearning.ai that had contributed to the materials 345 | 346 | 70 347 | 00:04:39,000 --> 00:04:42,000 348 | 他们为我和Isa将要介绍的资料做出了贡献 349 | that Isa and I will be presenting 350 | 351 | 71 352 | 00:04:42,000 --> 00:04:45,000 353 | 我非常感谢Andrew Main, Joe Palermo, Boris Power 354 | I'm very grateful to Andrew Main, Joe Palermo, Boris Power 355 | 356 | 72 357 | 00:04:45,000 --> 00:04:49,000 358 | Ted Sanders，以及来自OpenAI的Lilian Wang 359 | Ted Sanders, and Lilian Wang from OpenAI that were very involved 360 | 361 | 73 362 | 00:04:49,000 --> 00:04:53,000 363 | 与我们一起集思广益，审核资料 364 | with us brainstorming materials, vetting the materials to put together 365 | 366 | 74 367 | 00:04:53,000 --> 00:04:55,000 368 | 并做成这个简短的课程 369 | the curriculum for this short course 370 | 371 | 75 372 | 00:04:55,000 --> 00:04:58,000 373 | 我也很感谢DeepLearning方面的工作 374 | And I'm also grateful on the DeepLearning side for the work 375 | 376 | 76 377 | 00:04:58,000 --> 00:05:01,000 378 | Jeff Ludwig、Eddie Hsu和Tommy Nelson的工作 379 | of Jeff Ludwig, Eddie Hsu, and Tommy Nelson 380 | 381 | 77 382 | 00:05:01,000 --> 00:05:06,000 383 | 因此，当你使用一个指令学习LLM时，可以将其看作是 384 | So when you use an instruction-tuned LLM, think of giving instructions 385 | 386 | 78 387 | 00:05:06,000 --> 00:05:10,000 388 | 向另一个人发出指令，比如一个聪明但不知道 389 | to another person, say someone that's smart but doesn't know 390 | 391 | 79 392 | 00:05:10,000 --> 00:05:12,000 393 | 任务细节的人 394 | the specifics of your task 395 | 396 | 80 397 | 00:05:12,000 --> 00:05:16,000 398 | 因此，当一个LLM不工作时，有时是因为指令 399 | So when an LLM doesn't work, sometimes it's because the instructions 400 | 401 | 81 402 | 00:05:16,000 --> 00:05:17,000 403 | 不够清楚 404 | weren't clear enough 405 | 406 | 82 407 | 00:05:17,000 --> 00:05:20,000 408 | 例如，如果你说，请给我写一些 409 | For example, if you were to say, please write me something 410 | 411 | 83 412 | 00:05:20,000 --> 00:05:22,000 413 | 关于艾伦-图灵的东西 414 | about Alan Turing 415 | 416 | 84 417 | 00:05:22,000 --> 00:05:26,000 418 | 那么，除此之外，明确以下几点也会有所帮助 419 | Well, in addition to that, it can be helpful to be clear about 420 | 421 | 85 422 | 00:05:26,000 --> 00:05:30,000 423 | 你想让文章侧重于他的科学工作 424 | whether you want the text to focus on his scientific work 425 | 426 | 86 427 | 00:05:30,000 --> 00:05:34,000 428 | 他的个人生活、他在历史上的作用或其他方向 429 | or his personal life or his role in history or something else 430 | 431 | 87 432 | 00:05:34,000 --> 00:05:39,000 433 | 如果你指定了想要文本的语调 434 | And if you specify what you want the tone of the text to be 435 | 436 | 88 437 | 00:05:39,000 --> 00:05:43,000 438 | 它是否会根据指定的语调来写，比如像专业记者写的那样 439 | should it take on the tone like a professional journalist would write 440 | 441 | 89 442 | 00:05:43,000 --> 00:05:46,000 443 | 还是更像随手写给朋友的轻松语气？ 444 | or is it more of a casual note that you dash off to a friend? 445 | 446 | 90 447 | 00:05:46,000 --> 00:05:47,000 448 | 这一点是成立的 449 | That holds 450 | 451 | 91 452 | 00:05:47,000 --> 00:05:49,000 453 | LLM会产生你想要的东西 454 | The LLM generates what you want 455 | 456 | 92 457 | 00:05:49,000 --> 00:05:52,000 458 | 当然，如果你想象是自己在问，比如 459 | And of course, if you picture yourself asking, say 460 | 461 | 93 462 | 00:05:52,000 --> 00:05:56,000 463 | 一个刚毕业的大学生为你完成这项任务 464 | a fresh college graduate to carry out this task for you 465 | 466 | 94 467 | 00:05:56,000 --> 00:05:59,000 468 | 如果你甚至可以指定他们应该阅读哪些文本片段 469 | if you can even specify what snippets of text they should read 470 | 471 | 95 472 | 00:05:59,000 --> 00:06:02,000 473 | 提前写出这篇关于艾伦-图灵的文章 474 | in advance to write this text about Alan Turing 475 | 476 | 96 477 | 00:06:02,000 --> 00:06:06,000 478 | 那么，这能成功让那个刚毕业的大学生 479 | then that even better sets up that fresh college grad for success 480 | 481 | 97 482 | 00:06:06,000 --> 00:06:09,000 483 | 更好的为你完成这项任务 484 | to carry out this task for you 485 | 486 | 98 487 | 00:06:09,000 --> 00:06:13,000 488 | 因此，在接下来的视频中，你会看到如何清晰且明确 489 | So in the next video, you see examples of how to be clear 490 | 491 | 99 492 | 00:06:13,000 --> 00:06:17,000 493 | 描述提示，这是一个重要的LLM提示准则 494 | and specific, which is an important principle of prompting LLMs 495 | 496 | 100 497 | 00:06:17,000 --> 00:06:21,000 498 | 而且你还将从Isa那里学到了第二个提示词准则 499 | And you also learn from Isa a second principle of prompting 500 | 501 | 101 502 | 00:06:21,000 --> 00:06:24,000 503 | 那就是给LLM时间来思考 504 | that is giving a LLM time to think 505 | 506 | 102 507 | 00:06:24,000 --> 00:06:29,000 508 | 因此，让我们继续观看下一个视频 509 | So with that, let's go on to the next video 510 | 511 | -------------------------------------------------------------------------------- /chinese_subtitles/prompt_eng_05_inferring_chinese.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:07,080 3 | 接下来的这段视频是关于推断的 4 | 5 | 2 6 | 00:00:07,080 --> 00:00:09,540 7 | 我喜欢把这些任务看成是模型将一个文本作为输入并进行某种分析 8 | 9 | 3 10 | 00:00:09,540 --> 00:00:12,160 11 | 作为输入的文本并进行某种分析 12 | 13 | 4 14 | 00:00:12,160 --> 00:00:14,640 15 | 所以这可能是提取标签 16 | 17 | 5 18 | 00:00:14,640 --> 00:00:19,280 19 | 提取名字，理解文本的情感，诸如此类的事情 20 | 21 | 6 22 | 00:00:19,280 --> 00:00:22,480 23 | 因此，如果你想提取一个情感 24 | 25 | 7 26 | 00:00:22,480 --> 00:00:24,240 27 | 正面或负面的文本 28 | 29 | 8 30 | 00:00:24,240 --> 00:00:26,920 31 | 在传统的机器学习工作流程中 32 | 33 | 9 34 | 00:00:26,920 --> 00:00:29,680 35 | 你必须要收集标签数据集 36 | 37 | 10 38 | 00:00:29,680 --> 00:00:33,200 39 | 训练一个模型，弄清楚如何在云端的某个地方部署这个模型 40 | 41 | 11 42 | 00:00:33,200 --> 00:00:34,240 43 | 并做出推断 44 | 45 | 12 46 | 00:00:34,240 --> 00:00:35,320 47 | 这可以很好地工作 48 | 49 | 13 50 | 00:00:35,320 --> 00:00:38,560 51 | 但要经历这个过程，实在是一件很费力的事情 52 | 53 | 14 54 | 00:00:38,560 --> 00:00:43,680 55 | 此外，对于每一项任务，如情感分析与提取名字或其他东西 56 | 57 | 15 58 | 00:00:43,680 --> 00:00:46,920 59 | 你必须训练和部署一个单独的模型 60 | 61 | 16 62 | 00:00:46,920 --> 00:00:48,760 63 | 关于大语言模型的一个好处是 64 | 65 | 17 66 | 00:00:48,760 --> 00:00:51,400 67 | 对于许多像这样的任务 68 | 69 | 18 70 | 00:00:51,400 --> 00:00:56,240 71 | 你可以只写一个提示，然后让它马上开始生成结果 72 | 73 | 19 74 | 00:00:56,240 --> 00:01:00,080 75 | 这给应用程序的开发带来了巨大的提速 76 | 77 | 20 78 | 00:01:00,080 --> 00:01:02,360 79 | 你也可以只使用一个模型 80 | 81 | 21 82 | 00:01:02,360 --> 00:01:05,520 83 | 一个API来完成许多不同的任务，而不需要 84 | 85 | 22 86 | 00:01:05,520 --> 00:01:09,200 87 | 弄清楚如何训练和部署许多不同的模型 88 | 89 | 23 90 | 00:01:09,200 --> 00:01:13,840 91 | 有了这些，让我们跳进代码，看看你如何利用这个优势 92 | 93 | 24 94 | 00:01:13,840 --> 00:01:16,160 95 | 这里是我们通常的初始代码 96 | 97 | 25 98 | 00:01:16,160 --> 00:01:19,200 99 | 运行这个 100 | 101 | 26 102 | 00:01:19,200 --> 00:01:24,640 103 | 我使用的最多的例子，是关于一盏灯的评论 104 | 105 | 27 106 | 00:01:24,640 --> 00:01:26,880 107 | 卧室里需要一盏漂亮的灯 108 | 109 | 28 110 | 00:01:26,880 --> 00:01:30,240 111 | 和额外存储空间的家具，等等 112 | 113 | 29 114 | 00:01:31,920 --> 00:01:39,400 115 | 让我写一个提示，对这个的情绪进行分类 116 | 117 | 30 118 | 00:01:39,400 --> 00:01:44,520 119 | 如果我想让系统告诉我这是什么情绪 120 | 121 | 31 122 | 00:01:44,520 --> 00:01:55,720 123 | 我可以直接写出 "以下产品评论的情感是什么" 124 | 125 | 32 126 | 00:01:58,960 --> 00:02:04,520 127 | 加上通常的分隔符和评论文本等等，然后让我们运行这个提示 128 | 129 | 33 130 | 00:02:04,640 --> 00:02:08,880 131 | 这表示该产品评论的情绪是积极的 132 | 133 | 34 134 | 00:02:08,880 --> 00:02:11,400 135 | 这实际上似乎很正确 136 | 137 | 35 138 | 00:02:11,400 --> 00:02:14,720 139 | 这盏灯并不完美，但这位顾客看起来很满意 140 | 141 | 36 142 | 00:02:14,720 --> 00:02:17,000 143 | 这似乎是一个关心客户和产品的伟大公司 144 | 145 | 37 146 | 00:02:17,000 --> 00:02:20,360 147 | 我认为积极的情绪似乎是正确的答案 148 | 149 | 38 150 | 00:02:20,360 --> 00:02:22,440 151 | 现在，这就把整个句子打印出来 152 | 153 | 39 154 | 00:02:22,440 --> 00:02:25,480 155 | 该产品评论的情绪是积极的 156 | 157 | 40 158 | 00:02:25,480 --> 00:02:31,720 159 | 如果你想给出一个更简洁的回答，以便于后期处理 160 | 161 | 41 162 | 00:02:31,720 --> 00:02:35,120 163 | 我可以在这个提示中加入另一条指令 164 | 165 | 42 166 | 00:02:35,120 --> 00:02:37,320 167 | 只给你一个单词的答案 168 | 169 | 43 170 | 00:02:37,320 --> 00:02:38,960 171 | 正面或负面 172 | 173 | 44 174 | 00:02:38,960 --> 00:02:42,840 175 | 它只是像这样打印出正面，这对一段文字来说更容易 176 | 177 | 45 178 | 00:02:42,840 --> 00:02:46,760 179 | 来处理这个输出，并对它做一些处理 180 | 181 | 46 182 | 00:02:46,760 --> 00:02:49,320 183 | 让我们看一下另一个提示 184 | 185 | 47 186 | 00:02:49,320 --> 00:02:52,160 187 | 再来一遍，仍然使用关于台灯的评论 188 | 189 | 48 190 | 00:02:52,160 --> 00:02:55,760 191 | 在这里，我让它识别一个列表 192 | 193 | 49 194 | 00:02:55,760 --> 00:02:58,360 195 | 以下评论的作者所表达的情绪 196 | 197 | 50 198 | 00:02:58,360 --> 00:03:01,840 199 | 在这个列表中包括不超过五个项目 200 | 201 | 51 202 | 00:03:01,840 --> 00:03:04,680 203 | 大语言模型在以下方面相当出色 204 | 205 | 52 206 | 00:03:04,680 --> 00:03:08,320 207 | 从一段文本中提取特定的东西 208 | 209 | 53 210 | 00:03:08,320 --> 00:03:11,760 211 | 在这种情况下，我们要表达的是情绪 212 | 213 | 54 214 | 00:03:11,760 --> 00:03:14,720 215 | 这对于了解 216 | 217 | 55 218 | 00:03:14,720 --> 00:03:18,320 219 | 你的客户对某一特定产品的看法 220 | 221 | 56 222 | 00:03:18,320 --> 00:03:21,480 223 | 对于很多客户支持机构来说，了解客户对产品的看法很重要 224 | 225 | 57 226 | 00:03:21,480 --> 00:03:26,640 227 | 理解一个特定的用户是否对你的产品非常不满是非常重要的 228 | 229 | 58 230 | 00:03:26,640 --> 00:03:30,400 231 | 你可能有这样一个不同的分类问题 232 | 233 | 59 234 | 00:03:30,400 --> 00:03:33,040 235 | 下面这篇评论的作者是否在表达愤怒？ 236 | 237 | 60 238 | 00:03:33,040 --> 00:03:34,880 239 | 因为如果某人真的很生气 240 | 241 | 61 242 | 00:03:34,880 --> 00:03:39,840 243 | 它可能值得付出额外的关注来获得客户评论 244 | 245 | 62 246 | 00:03:39,840 --> 00:03:41,920 247 | 获得客户支持或成功拿下客户 248 | 249 | 63 250 | 00:03:41,920 --> 00:03:46,000 251 | 伸出援手，弄清楚发生了什么事，并为客户把事情做好 252 | 253 | 64 254 | 00:03:46,000 --> 00:03:48,960 255 | 在这种情况下，客户并不生气 256 | 257 | 65 258 | 00:03:48,960 --> 00:03:52,240 259 | 注意，有了监督学习 260 | 261 | 66 262 | 00:03:52,240 --> 00:03:55,400 263 | 如果我想建立所有这些分类器 264 | 265 | 67 266 | 00:03:55,400 --> 00:03:58,880 267 | 我不可能用监督学习在短短几分钟内完成这些工作 268 | 269 | 68 270 | 00:03:58,880 --> 00:04:04,040 271 | 而你在这段视频中看到我这样做了 272 | 273 | 69 274 | 00:04:04,040 --> 00:04:08,520 275 | 我鼓励你暂停这个视频，并尝试改变其中的一些提示 276 | 277 | 70 278 | 00:04:08,520 --> 00:04:12,680 279 | 也许问一下客户是否表达了喜悦之情，或者问一下是否有 280 | 281 | 71 282 | 00:04:12,680 --> 00:04:15,320 283 | 任何缺失的部分，看看你是否能得到一个提示来 284 | 285 | 72 286 | 00:04:15,320 --> 00:04:20,120 287 | 对这个实验室进行不同的推断 288 | 289 | 73 290 | 00:04:20,600 --> 00:04:28,360 291 | 让我再展示一下你可以用这个系统做的一些事情 292 | 293 | 74 294 | 00:04:28,360 --> 00:04:35,360 295 | 特别是从客户评论中提取更丰富的信息 296 | 297 | 75 298 | 00:04:35,640 --> 00:04:39,920 299 | 信息提取是NLP的一部分 300 | 301 | 76 302 | 00:04:39,920 --> 00:04:43,160 303 | 自然语言处理的一部分，它涉及到将一段文本 304 | 305 | 77 306 | 00:04:43,160 --> 00:04:48,240 307 | 并从文本中提取出你想知道的某些东西 308 | 309 | 78 310 | 00:04:48,240 --> 00:04:50,600 311 | 在这个提示中，我问它 312 | 313 | 79 314 | 00:04:50,600 --> 00:04:52,840 315 | 确定以下项目 316 | 317 | 80 318 | 00:04:52,840 --> 00:04:57,360 319 | 购买的物品和制造该物品的公司名称 320 | 321 | 81 322 | 00:04:57,360 --> 00:05:05,320 323 | 同样，如果你想从一个在线购物的电子商务网站上总结出许多评论 324 | 325 | 82 326 | 00:05:05,320 --> 00:05:07,920 327 | 对于你收集的大量评论来说 328 | 329 | 83 330 | 00:05:07,920 --> 00:05:10,160 331 | 弄清楚这些评论所涉及的物品可能会很有用 332 | 333 | 84 334 | 00:05:10,160 --> 00:05:13,600 335 | 可以通过分析评论中的内容，找出评论涉及的物品制造商 336 | 337 | 85 338 | 00:05:13,600 --> 00:05:16,680 339 | 并确定积极和消极情绪 340 | 341 | 86 342 | 00:05:16,680 --> 00:05:20,480 343 | 以追踪特定物品或制造商的积极或消极情绪趋势 344 | 345 | 87 346 | 00:05:20,480 --> 00:05:24,960 347 | 在这个例子中，我将要求它把响应格式化为 348 | 349 | 88 350 | 00:05:24,960 --> 00:05:29,560 351 | 一个以项目和品牌为键的JSON对象 352 | 353 | 89 354 | 00:05:29,560 --> 00:05:33,840 355 | 如果我这样做了，它就会说这个物品是一盏灯 356 | 357 | 90 358 | 00:05:33,840 --> 00:05:35,880 359 | 品牌是Luminar 360 | 361 | 91 362 | 00:05:35,880 --> 00:05:38,000 363 | 你可以很容易地将其加载到 364 | 365 | 92 366 | 00:05:38,000 --> 00:05:42,880 367 | Python字典中，然后对这个输出做额外的处理 368 | 369 | 93 370 | 00:05:42,880 --> 00:05:44,600 371 | 在我们所经历的例子中 372 | 373 | 94 374 | 00:05:44,600 --> 00:05:48,640 375 | 你看到了如何写一个提示来识别情感 376 | 377 | 95 378 | 00:05:48,640 --> 00:05:50,960 379 | 弄清某人是否生气 380 | 381 | 96 382 | 00:05:50,960 --> 00:05:54,600 383 | 然后再提取物品和品牌 384 | 385 | 97 386 | 00:05:54,600 --> 00:06:01,920 387 | 提取所有这些信息的一个方法是使用三或四个提示语 388 | 389 | 98 390 | 00:06:01,920 --> 00:06:06,400 391 | 并调用三次或四次get_completion 392 | 393 | 99 394 | 00:06:06,400 --> 00:06:09,960 395 | 一次一次地提取这些不同的字段 396 | 397 | 100 398 | 00:06:09,960 --> 00:06:13,080 399 | 但事实证明，你实际上可以只写一个提示符 400 | 401 | 101 402 | 00:06:13,080 --> 00:06:16,480 403 | 来同时提取所有这些信息 404 | 405 | 102 406 | 00:06:16,480 --> 00:06:18,960 407 | 比方说识别以下项目 408 | 409 | 103 410 | 00:06:18,960 --> 00:06:23,040 411 | 提取情绪，是评论者表达愤怒 412 | 413 | 104 414 | 00:06:23,040 --> 00:06:25,840 415 | 项目的购买，公司的动画 416 | 417 | 105 418 | 00:06:25,840 --> 00:06:33,480 419 | 然后在这里，我还要告诉它把愤怒值格式化为一个布尔值 420 | 421 | 106 422 | 00:06:33,480 --> 00:06:35,960 423 | 让我运行它 424 | 425 | 107 426 | 00:06:35,960 --> 00:06:43,480 427 | 这将输出一个JSON，其中情绪是积极的 428 | 429 | 108 430 | 00:06:43,480 --> 00:06:45,880 431 | 愤怒，然后在false周围没有引号 432 | 433 | 109 434 | 00:06:45,880 --> 00:06:48,840 435 | 因为它也只是把它输出为一个布尔值 436 | 437 | 110 438 | 00:06:48,840 --> 00:06:52,520 439 | 提取出"lamp with additional storage"作为物品，而不仅仅是"lamp" 440 | 441 | 111 442 | 00:06:52,520 --> 00:06:53,880 443 | 看起来还不错 444 | 445 | 112 446 | 00:06:53,880 --> 00:06:57,760 447 | 这种方法让你只需一个提示 448 | 449 | 113 450 | 00:06:57,760 --> 00:07:02,400 451 | 就可以从一段文本中提取多个字段 452 | 453 | 114 454 | 00:07:02,400 --> 00:07:05,200 455 | 像往常一样，请随意暂停视频 456 | 457 | 115 458 | 00:07:05,200 --> 00:07:08,160 459 | 并在此基础上试玩一下，感受不同变化 460 | 461 | 116 462 | 00:07:08,160 --> 00:07:12,080 463 | 或者甚至可以尝试输入一个完全不同的评论，看看你是否 464 | 465 | 117 466 | 00:07:12,080 --> 00:07:15,480 467 | 能否仍然准确地提取这些东西 468 | 469 | 118 470 | 00:07:15,480 --> 00:07:19,440 471 | 现在，我看到大型语言模型很酷的应用之一 472 | 473 | 119 474 | 00:07:19,440 --> 00:07:22,600 475 | 是推断主题 476 | 477 | 120 478 | 00:07:22,600 --> 00:07:25,000 479 | 给出一段长的文字 480 | 481 | 121 482 | 00:07:25,000 --> 00:07:28,040 483 | 这段文字是关于什么的？ 484 | 485 | 122 486 | 00:07:28,040 --> 00:07:29,480 487 | 主题是什么？ 488 | 489 | 123 490 | 00:07:29,480 --> 00:07:33,360 491 | 这是一篇虚构的报纸文章，关于如何 492 | 493 | 124 494 | 00:07:33,360 --> 00:07:37,720 495 | 政府工作人员对他们所工作的机构的感受 496 | 497 | 125 498 | 00:07:37,720 --> 00:07:41,440 499 | 最近由政府等进行的调查 500 | 501 | 126 502 | 00:07:41,440 --> 00:07:43,160 503 | 结果审查，NASA是 504 | 505 | 127 506 | 00:07:43,160 --> 00:07:46,400 507 | 一个受欢迎的部门，满意度很高 508 | 509 | 128 510 | 00:07:46,400 --> 00:07:48,000 511 | 我是NASA的粉丝 512 | 513 | 129 514 | 00:07:48,000 --> 00:07:51,040 515 | 喜欢他们所做的工作，但这是一篇虚构的文章 516 | 517 | 130 518 | 00:07:51,040 --> 00:07:53,280 519 | 鉴于这样的文章 520 | 521 | 131 522 | 00:07:53,280 --> 00:07:57,920 523 | 我们可以用这个提示来问它 524 | 525 | 132 526 | 00:07:57,920 --> 00:08:03,320 527 | 确定以下文本中讨论的五个主题 528 | 529 | 133 530 | 00:08:03,320 --> 00:08:05,760 531 | 让我们把每一项都写成一到两个字的长度 532 | 533 | 134 534 | 00:08:05,760 --> 00:08:09,320 535 | 为我的回答，用逗号分隔的列表 536 | 537 | 135 538 | 00:08:09,320 --> 00:08:11,160 539 | 如果我们运行这个 540 | 541 | 136 542 | 00:08:11,160 --> 00:08:14,680 543 | 我们就会发现，这篇文章是关于政府调查的 544 | 545 | 137 546 | 00:08:14,680 --> 00:08:17,880 547 | 是关于工作满意度的，是关于NASA的，等等 548 | 549 | 138 550 | 00:08:17,880 --> 00:08:23,320 551 | 总的来说，我认为相当不错的主题列表的提取 552 | 553 | 139 554 | 00:08:23,320 --> 00:08:28,280 555 | 当然，你也可以把它分开，这样你就可以得到 556 | 557 | 140 558 | 00:08:28,280 --> 00:08:34,360 559 | 一个包含这篇文章所涉及的五个主题的Python列表 560 | 561 | 141 562 | 00:08:34,360 --> 00:08:38,600 563 | 如果你有一个文章的集合，并提取主题 564 | 565 | 142 566 | 00:08:38,600 --> 00:08:40,800 567 | 那么你还可以使用 568 | 569 | 143 570 | 00:08:40,800 --> 00:08:45,640 571 | 一个大语言模型来帮助你索引到不同的主题 572 | 573 | 144 574 | 00:08:45,640 --> 00:08:48,560 575 | 让我用一个稍微不同的主题列表 576 | 577 | 145 578 | 00:08:48,560 --> 00:08:52,240 579 | 比方说，我们是一个新闻网站或其他什么 580 | 581 | 146 582 | 00:08:52,240 --> 00:08:54,720 583 | 而这些是我们追踪的主题 584 | 585 | 147 586 | 00:08:54,720 --> 00:08:56,480 587 | NASA，当地政府，工程 588 | 589 | 148 590 | 00:08:56,480 --> 00:08:59,480 591 | 员工满意度，联邦政府 592 | 593 | 149 594 | 00:08:59,480 --> 00:09:01,680 595 | 比方说，你想弄清楚 596 | 597 | 150 598 | 00:09:01,680 --> 00:09:03,280 599 | 给出一篇新闻文章 600 | 601 | 151 602 | 00:09:03,280 --> 00:09:07,720 603 | 这篇新闻报道中涉及哪些主题 604 | 605 | 152 606 | 00:09:07,720 --> 00:09:10,760 607 | 这里有一个提示，我可以使用 608 | 609 | 153 610 | 00:09:10,760 --> 00:09:13,720 611 | 我想说的是，确定 "有趣的主题清单 "中的每一项 612 | 613 | 154 614 | 00:09:13,720 --> 00:09:16,840 615 | 是否是下面文本中的一个主题 616 | 617 | 155 618 | 00:09:16,840 --> 00:09:20,880 619 | 给出你的答案，这是每个话题的01 620 | 621 | 156 622 | 00:09:22,320 --> 00:09:25,880 623 | 很好。这是和之前一样的故事文本 624 | 625 | 157 626 | 00:09:25,880 --> 00:09:28,240 627 | 这个东西是一个故事。它是关于NASA的 628 | 629 | 158 630 | 00:09:28,240 --> 00:09:30,840 631 | 这不是关于地方政府，不是关于工程 632 | 633 | 159 634 | 00:09:30,840 --> 00:09:35,080 635 | 它是关于员工满意度的，它是关于联邦政府的 636 | 637 | 160 638 | 00:09:35,080 --> 00:09:37,720 639 | 有了这个，在机器学习中 640 | 641 | 161 642 | 00:09:37,720 --> 00:09:42,320 643 | 这有时被称为 "零示例学习算法" 644 | 645 | 162 646 | 00:09:42,320 --> 00:09:45,600 647 | 因为我们没有给它任何有标签的训练数据 648 | 649 | 163 650 | 00:09:45,600 --> 00:09:47,440 651 | 这就是零示例 652 | 653 | 164 654 | 00:09:47,440 --> 00:09:48,960 655 | 只需一个提示 656 | 657 | 165 658 | 00:09:48,960 --> 00:09:51,480 659 | 它就能够确定哪些 660 | 661 | 166 662 | 00:09:51,480 --> 00:09:55,880 663 | 这些主题在那篇新闻报道中被涵盖 664 | 665 | 167 666 | 00:09:55,880 --> 00:10:00,800 667 | 如果你想生成一个新闻警报 668 | 669 | 168 670 | 00:10:00,800 --> 00:10:05,840 671 | 这样处理新闻，我真的很喜欢NASA做的很多工作 672 | 673 | 169 674 | 00:10:05,840 --> 00:10:10,560 675 | 如果你想建立一个系统，可以把这个 676 | 677 | 170 678 | 00:10:10,560 --> 00:10:13,440 679 | 把这些信息放到一个字典里，每当 680 | 681 | 171 682 | 00:10:13,440 --> 00:10:15,960 683 | NASA的新闻出现，就打印提醒 684 | 685 | 172 686 | 00:10:15,960 --> 00:10:20,520 687 | 新的NASA故事，他们可以用这个来非常迅速地采取任何文章 688 | 689 | 173 690 | 00:10:20,520 --> 00:10:22,680 691 | 弄清楚它是关于什么主题的 692 | 693 | 174 694 | 00:10:22,680 --> 00:10:24,600 695 | 如果该主题包括NASA 696 | 697 | 175 698 | 00:10:24,600 --> 00:10:27,920 699 | 让它打印出警报，新的NASA故事 700 | 701 | 176 702 | 00:10:27,920 --> 00:10:32,480 703 | 只有一件事，我在下面使用这个主题词典 704 | 705 | 177 706 | 00:10:32,480 --> 00:10:36,200 707 | 我在上面使用的这个提示不是很强大 708 | 709 | 178 710 | 00:10:36,200 --> 00:10:38,080 711 | 如果我想要一个生产系统 712 | 713 | 179 714 | 00:10:38,080 --> 00:10:43,680 715 | 我可能会让它以JSON格式输出答案 716 | 717 | 180 718 | 00:10:43,680 --> 00:10:46,400 719 | 而不是作为一个列表，因为输出的 720 | 721 | 181 722 | 00:10:46,400 --> 00:10:49,640 723 | 大语言模型会有一点不一致 724 | 725 | 182 726 | 00:10:49,640 --> 00:10:52,240 727 | 这实际上是一段很脆弱 的代码 728 | 729 | 183 730 | 00:10:52,240 --> 00:10:54,800 731 | 但是，如果你想，当你看完这个视频后 732 | 733 | 184 734 | 00:10:54,800 --> 00:10:58,520 735 | 可以看看你能不能想出办法来修改这个提示 736 | 737 | 185 738 | 00:10:58,520 --> 00:11:01,320 739 | 让它输出JSON而不是像这样的一个列表 740 | 741 | 186 742 | 00:11:01,320 --> 00:11:07,720 743 | 然后有一个更强大的方法来判断某篇文章是否是关于NASA的故事 744 | 745 | 187 746 | 00:11:07,720 --> 00:11:10,800 747 | 所以，这就是推断的内容 748 | 749 | 188 750 | 00:11:10,800 --> 00:11:12,600 751 | 在短短的几分钟内 752 | 753 | 189 754 | 00:11:12,600 --> 00:11:16,720 755 | 你可以建立多个系统，对文本进行推断 756 | 757 | 190 758 | 00:11:16,720 --> 00:11:19,760 759 | 这些推断以前需要一名熟练的机器学习开发人员 760 | 761 | 191 762 | 00:11:19,760 --> 00:11:23,680 763 | 花费数天甚至数周才能完成 764 | 765 | 192 766 | 00:11:23,680 --> 00:11:27,120 767 | 我认为，无论是对于熟练的机器学习开发人员 768 | 769 | 193 770 | 00:11:27,120 --> 00:11:29,280 771 | 还是机器学习新手而言 772 | 773 | 194 774 | 00:11:29,280 --> 00:11:31,720 775 | 这都是非常令人兴奋的消息 776 | 777 | 195 778 | 00:11:31,720 --> 00:11:37,080 779 | 你现在可以使用提示来非常快速地建立并开始 780 | 781 | 196 782 | 00:11:37,080 --> 00:11:42,880 783 | 对这些相当复杂的自然语言处理任务做出推断 784 | 785 | 197 786 | 00:11:42,880 --> 00:11:44,240 787 | 在下一个视频中 788 | 789 | 198 790 | 00:11:44,240 --> 00:11:47,520 791 | 我们将继续讨论令人激动的事情，你可以用 792 | 793 | 199 794 | 00:11:47,520 --> 00:11:51,240 795 | 大语言模型，我们将继续进行转换 796 | 797 | 200 798 | 00:11:51,240 --> 00:11:54,800 799 | 把一段文本转换为 800 | 801 | 201 802 | 00:11:54,800 --> 00:11:58,440 803 | 不同的文本，如翻译成不同的语言 804 | 805 | 202 806 | 00:11:58,440 --> 00:12:07,600 807 | 让我们继续看下一个视频 808 | 809 | -------------------------------------------------------------------------------- /chinese_subtitles/prompt_eng_08_chatbot_chinese.srt: -------------------------------------------------------------------------------- 1 | 1 2 | 00:00:05,000 --> 00:00:08,600 3 | 关于大语言模型的一个令人兴奋的事情，是你可以用它来 4 | 5 | 2 6 | 00:00:08,600 --> 00:00:12,680 7 | 构建一个自定义的聊天机器人，只需花费少量的精力 8 | 9 | 3 10 | 00:00:12,680 --> 00:00:17,660 11 | ChatGPT的Web界面能让你通过一个大语言模型 12 | 13 | 4 14 | 00:00:17,660 --> 00:00:20,500 15 | 进行对话 16 | 17 | 5 18 | 00:00:20,500 --> 00:00:23,120 19 | 但其中一个很酷的事情是你也可以使用 20 | 21 | 6 22 | 00:00:23,120 --> 00:00:25,680 23 | 一个大语言模型来建立你的自定义聊天机器人 24 | 25 | 7 26 | 00:00:25,680 --> 00:00:28,920 27 | 也许可以扮演一个AI客户服务代理或 28 | 29 | 8 30 | 00:00:28,920 --> 00:00:31,480 31 | 餐馆的AI接单员 32 | 33 | 9 34 | 00:00:31,480 --> 00:00:34,400 35 | 在这段视频中，你将学习如何做这件事 36 | 37 | 10 38 | 00:00:34,400 --> 00:00:37,080 39 | 我将描述以下内容 40 | 41 | 11 42 | 00:00:37,080 --> 00:00:39,960 43 | 我将更详细地描述OpenAI的聊天补全格式 44 | 45 | 12 46 | 00:00:39,960 --> 00:00:42,540 47 | 然后你将自己建立一个聊天机器人 48 | 49 | 13 50 | 00:00:42,540 --> 00:00:44,400 51 | 所以让我们开始吧 52 | 53 | 14 54 | 00:00:44,400 --> 00:00:49,000 55 | 首先，我们将像往常一样设置OpenAI Python包 56 | 57 | 15 58 | 00:00:49,000 --> 00:00:53,380 59 | 因此，像ChatGPT这样的聊天模型实际上是被训练成以一系列的信息作为输入 60 | 61 | 16 62 | 00:00:53,380 --> 00:00:57,640 63 | 并返回模型生成的信息作为输出 64 | 65 | 17 66 | 00:00:57,640 --> 00:00:59,960 67 | 因此，尽管聊天格式的设计是为了 68 | 69 | 18 70 | 00:00:59,960 --> 00:01:02,320 71 | 使得像这样的多轮对话变得容易 72 | 73 | 19 74 | 00:01:02,320 --> 00:01:06,000 75 | 我们在之前的视频中已经看到 76 | 77 | 20 78 | 00:01:06,000 --> 00:01:09,560 79 | 它也适用于没有对话的单个任务 80 | 81 | 21 82 | 00:01:09,560 --> 00:01:14,200 83 | 所以接下来，我们要定义两个辅助函数 84 | 85 | 22 86 | 00:01:14,200 --> 00:01:17,080 87 | 就是我们在所有视频中一直在使用的一个 88 | 89 | 23 90 | 00:01:17,080 --> 00:01:20,360 91 | get_completion函数 92 | 93 | 24 94 | 00:01:20,360 --> 00:01:22,080 95 | 但是，如果你看一下它 96 | 97 | 25 98 | 00:01:22,080 --> 00:01:25,040 99 | 我们给出了一个提示，但是在这个函数里面 100 | 101 | 26 102 | 00:01:25,040 --> 00:01:29,320 103 | 我们实际上在做的是把这个提示放到一个 104 | 105 | 27 106 | 00:01:29,320 --> 00:01:32,160 107 | 看起来像某种用户信息 108 | 109 | 28 110 | 00:01:32,160 --> 00:01:37,200 111 | 这是因为ChatGPT模型是一个聊天模型 112 | 113 | 29 114 | 00:01:37,200 --> 00:01:40,200 115 | 这意味着它被训练成以一系列信息作为输入 116 | 117 | 30 118 | 00:01:40,200 --> 00:01:44,000 119 | 然后返回模型生成的信息输出 120 | 121 | 31 122 | 00:01:44,000 --> 00:01:45,880 123 | 所以用户信息是一种输入 124 | 125 | 32 126 | 00:01:45,880 --> 00:01:49,320 127 | 然后助理的信息是输出 128 | 129 | 33 130 | 00:01:49,320 --> 00:01:51,320 131 | 所以在这个视频中 132 | 133 | 34 134 | 00:01:51,320 --> 00:01:53,760 135 | 我们将使用一个不同的辅助函数 136 | 137 | 35 138 | 00:01:53,760 --> 00:01:55,640 139 | 而不是把 140 | 141 | 36 142 | 00:01:55,640 --> 00:01:58,400 143 | 单一提示作为输入并获得单一结果 144 | 145 | 37 146 | 00:01:58,400 --> 00:02:01,680 147 | 我们将传入一个信息列表 148 | 149 | 38 150 | 00:02:01,680 --> 00:02:05,400 151 | 这些信息可以来自各种不同的角色 152 | 153 | 39 154 | 00:02:05,400 --> 00:02:07,160 155 | 所以我将描述这些 156 | 157 | 40 158 | 00:02:07,160 --> 00:02:10,040 159 | 这里有一个消息列表的例子 160 | 161 | 41 162 | 00:02:10,040 --> 00:02:12,040 163 | 第一条信息是 164 | 165 | 42 166 | 00:02:12,040 --> 00:02:15,200 167 | 一个系统消息，它给出了一个总体指令 168 | 169 | 43 170 | 00:02:15,200 --> 00:02:16,360 171 | 然后在这个消息之后 172 | 173 | 44 174 | 00:02:16,360 --> 00:02:20,040 175 | 我们在用户和助理之间有几轮对话 176 | 177 | 45 178 | 00:02:20,040 --> 00:02:21,920 179 | 而且这种对话通常会继续下去 180 | 181 | 46 182 | 00:02:21,920 --> 00:02:24,800 183 | 如果你曾经使用过ChatGPT的Web界面 184 | 185 | 47 186 | 00:02:24,800 --> 00:02:27,680 187 | 那么你输入的信息就是用户信息 188 | 189 | 48 190 | 00:02:27,680 --> 00:02:30,880 191 | 然后ChatGPT的信息是助理的信息 192 | 193 | 49 194 | 00:02:30,880 --> 00:02:33,440 195 | 因此，系统信息有助于在某种程度上 196 | 197 | 50 198 | 00:02:33,440 --> 00:02:35,800 199 | 设定助理的行为和角色 200 | 201 | 51 202 | 00:02:35,800 --> 00:02:38,920 203 | 它就像对话的高级指令一样 204 | 205 | 52 206 | 00:02:38,920 --> 00:02:40,640 207 | 所以你可以把它看成是 208 | 209 | 53 210 | 00:02:40,640 --> 00:02:42,440 211 | 在助理的耳边说的话，并引导 212 | 213 | 54 214 | 00:02:42,440 --> 00:02:46,320 215 | 它的反应，而用户并不知道系统的信息 216 | 217 | 55 218 | 00:02:46,320 --> 00:02:47,960 219 | 因此，作为用户 220 | 221 | 56 222 | 00:02:47,960 --> 00:02:49,400 223 | 如果你曾经使用过ChatGPT 224 | 225 | 57 226 | 00:02:49,400 --> 00:02:52,640 227 | 你可能不知道ChatGPT的系统消息里有什么 228 | 229 | 58 230 | 00:02:52,640 --> 00:02:55,360 231 | 系统消息的好处是 232 | 233 | 59 234 | 00:02:55,360 --> 00:02:57,240 235 | 它为开发者提供了一种方法来框定会话 236 | 237 | 60 238 | 00:02:57,240 --> 00:02:59,120 239 | 而不是让请求本身 240 | 241 | 61 242 | 00:02:59,120 --> 00:03:01,320 243 | 成为对话的一部分 244 | 245 | 62 246 | 00:03:01,320 --> 00:03:04,880 247 | 这样，你就可以在不让用户察觉的情况下引导助理 248 | 249 | 63 250 | 00:03:04,880 --> 00:03:09,640 251 | 并指导其回复 252 | 253 | 64 254 | 00:03:10,280 --> 00:03:15,440 255 | 所以现在让我们试着在对话中使用这些信息 256 | 257 | 65 258 | 00:03:16,080 --> 00:03:19,160 259 | 所以我们将使用我们的新助理函数 260 | 261 | 66 262 | 00:03:19,160 --> 00:03:21,720 263 | 来补全消息内容 264 | 265 | 67 266 | 00:03:22,600 --> 00:03:27,200 267 | 而且我们也在使用更高的temperature 268 | 269 | 68 270 | 00:03:27,200 --> 00:03:29,400 271 | 所以系统消息说 272 | 273 | 69 274 | 00:03:29,400 --> 00:03:31,440 275 | 你是一个说话像莎士比亚的助理 276 | 277 | 70 278 | 00:03:31,440 --> 00:03:33,160 279 | 所以这是我们在向助理描述 280 | 281 | 71 282 | 00:03:33,160 --> 00:03:35,960 283 | 向助理描述它应该如何表现 284 | 285 | 72 286 | 00:03:35,960 --> 00:03:38,680 287 | 然后第一个用户信息是，告诉我一个笑话 288 | 289 | 73 290 | 00:03:38,680 --> 00:03:41,640 291 | 接下来是，鸡为什么要过马路？ 292 | 293 | 74 294 | 00:03:41,640 --> 00:03:45,520 295 | 然后最后的用户信息是，我不知道 296 | 297 | 75 298 | 00:03:45,520 --> 00:03:49,880 299 | 所以，如果我们运行这个 300 | 301 | 76 302 | 00:03:50,600 --> 00:03:55,400 303 | 响应是"去到另一边"。让我们再试一次 304 | 305 | 77 306 | 00:03:55,720 --> 00:03:57,920 307 | "要到另一边去" 308 | 309 | 78 310 | 00:03:57,920 --> 00:03:59,240 311 | 公平的先生或女士 312 | 313 | 79 314 | 00:03:59,240 --> 00:04:02,040 315 | 这是一个古老而经典的方法，永远不会失败 316 | 317 | 80 318 | 00:04:02,040 --> 00:04:05,120 319 | 所以这就是我们的莎士比亚式响应 320 | 321 | 81 322 | 00:04:05,640 --> 00:04:08,960 323 | 让我们再试一次 324 | 325 | 82 326 | 00:04:08,960 --> 00:04:11,520 327 | 因为我想让它更清楚，这是助理的信息 328 | 329 | 83 330 | 00:04:11,520 --> 00:04:19,360 331 | 所以在这里，让我们去打印整个消息的回应 332 | 333 | 84 334 | 00:04:32,880 --> 00:04:36,240 335 | 所以只是为了让这个更清楚 336 | 337 | 85 338 | 00:04:36,240 --> 00:04:41,160 339 | 这个响应是一个助理信息 340 | 341 | 86 342 | 00:04:41,160 --> 00:04:45,680 343 | 所以角色是助理，然后内容是消息本身 344 | 345 | 87 346 | 00:04:45,680 --> 00:04:48,000 347 | 所以这就是在这个辅助函数中发生的事情 348 | 349 | 88 350 | 00:04:48,000 --> 00:04:51,880 351 | 我们只是传出了消息的内容 352 | 353 | 89 354 | 00:05:04,960 --> 00:05:07,960 355 | 所以现在让我们再做一个例子 356 | 357 | 90 358 | 00:05:07,960 --> 00:05:09,840 359 | 这里我们的消息是 360 | 361 | 91 362 | 00:05:09,840 --> 00:05:13,160 363 | 系统信息是你是一个友好的聊天机器人 364 | 365 | 92 366 | 00:05:13,160 --> 00:05:14,560 367 | 而第一个用户信息是 368 | 369 | 93 370 | 00:05:14,560 --> 00:05:16,120 371 | 嗨，我的名字是Isa 372 | 373 | 94 374 | 00:05:16,120 --> 00:05:20,640 375 | 我们想获得第一个用户的信息 376 | 377 | 95 378 | 00:05:20,640 --> 00:05:22,080 379 | 所以让我们执行这个 380 | 381 | 96 382 | 00:05:22,080 --> 00:05:24,640 383 | 第一条助理信息 384 | 385 | 97 386 | 00:05:24,640 --> 00:05:26,880 387 | 因此，第一条信息是 388 | 389 | 98 390 | 00:05:26,880 --> 00:05:28,400 391 | 你好，Isa，很高兴见到你 392 | 393 | 99 394 | 00:05:28,400 --> 00:05:30,760 395 | 我今天有什么可以帮助你的吗？ 396 | 397 | 100 398 | 00:05:36,000 --> 00:05:38,760 399 | 现在让我们试试另一个例子 400 | 401 | 101 402 | 00:05:39,320 --> 00:05:43,720 403 | 所以这里我们的信息是系统信息 404 | 405 | 102 406 | 00:05:43,720 --> 00:05:45,120 407 | 你是一个友好的聊天机器人 408 | 409 | 103 410 | 00:05:45,120 --> 00:05:47,080 411 | 和第一个用户信息是 412 | 413 | 104 414 | 00:05:47,080 --> 00:05:53,880 415 | 是的，你能提醒我我的名字是什么吗？ 416 | 417 | 105 418 | 00:05:53,880 --> 00:05:56,400 419 | 让我们得到回应 420 | 421 | 106 422 | 00:05:56,400 --> 00:06:01,440 423 | 正如你所看到的 424 | 425 | 107 426 | 00:06:01,440 --> 00:06:05,800 427 | 该模型实际上并不知道我的名字 428 | 429 | 108 430 | 00:06:05,800 --> 00:06:10,680 431 | 因此，与语言模型的每一次对话都是一个独立的互动 432 | 433 | 109 434 | 00:06:10,680 --> 00:06:13,200 435 | 这意味着，你必须提供所有相关的信息 436 | 437 | 110 438 | 00:06:13,200 --> 00:06:16,120 439 | 供模型在当前对话中使用 440 | 441 | 111 442 | 00:06:16,120 --> 00:06:18,560 443 | 如果你想让模型从或中汲取 444 | 445 | 112 446 | 00:06:18,560 --> 00:06:21,240 447 | 引述，记住一个对话的早期部分 448 | 449 | 113 450 | 00:06:21,240 --> 00:06:24,640 451 | 你必须在模型的输入中提供早期的对话交互内容 452 | 453 | 114 454 | 00:06:24,640 --> 00:06:26,960 455 | 我们也把这称为上下文 456 | 457 | 115 458 | 00:06:26,960 --> 00:06:30,240 459 | 所以，让我们来试试这个 460 | 461 | 116 462 | 00:06:34,440 --> 00:06:39,000 463 | 所以现在我们已经给出了模型需要的上下文 464 | 465 | 117 466 | 00:06:39,000 --> 00:06:41,600 467 | 这是我在以前的信息中的名字 468 | 469 | 118 470 | 00:06:41,600 --> 00:06:43,160 471 | 我们将问同样的问题 472 | 473 | 119 474 | 00:06:43,160 --> 00:06:45,640 475 | 所以我们会问我的名字是什么 476 | 477 | 120 478 | 00:06:45,960 --> 00:06:50,320 479 | 模型能够作出反应，因为它有所有的上下文 480 | 481 | 121 482 | 00:06:50,320 --> 00:06:55,520 483 | 在我们输入的消息列表中需要它 484 | 485 | 122 486 | 00:06:55,520 --> 00:06:58,760 487 | 所以现在你要建立你自己的聊天机器人 488 | 489 | 123 490 | 00:06:58,760 --> 00:07:02,040 491 | 这个聊天机器人将被称为orderbot（点餐机器人） 492 | 493 | 124 494 | 00:07:02,040 --> 00:07:05,520 495 | 我们将自动收集 496 | 497 | 125 498 | 00:07:05,520 --> 00:07:09,480 499 | 我们将自动收集用户的提示和助理的回应，以建立这个订单机器人 500 | 501 | 126 502 | 00:07:09,480 --> 00:07:12,200 503 | 它将在一家披萨店接受订单 504 | 505 | 127 506 | 00:07:12,200 --> 00:07:16,040 507 | 所以首先，我们要定义这个助理函数 508 | 509 | 128 510 | 00:07:16,040 --> 00:07:19,040 511 | 这个函数要做的是收集 512 | 513 | 129 514 | 00:07:19,040 --> 00:07:24,840 515 | 我们的用户信息，这样我们就可以避免像上面那样用手输入信息了 516 | 517 | 130 518 | 00:07:24,840 --> 00:07:29,760 519 | 这将从我们下面建立的用户界面上收集提示信息 520 | 521 | 131 522 | 00:07:29,760 --> 00:07:32,800 523 | 然后将其追加到一个叫做"上下文"的列表中 524 | 525 | 132 526 | 00:07:32,800 --> 00:07:36,280 527 | 然后它每次都会用这个上下文调用模型 528 | 529 | 133 530 | 00:07:36,280 --> 00:07:39,320 531 | 然后模型的响应也会被添加到上下文中 532 | 533 | 134 534 | 00:07:39,320 --> 00:07:43,560 535 | 所以模型的信息被添加到上下文中 536 | 537 | 135 538 | 00:07:43,560 --> 00:07:45,160 539 | 用户信息被添加到上下文中 540 | 541 | 136 542 | 00:07:45,160 --> 00:07:47,880 543 | 以此类推，它只会越来越长 544 | 545 | 137 546 | 00:07:47,880 --> 00:07:52,680 547 | 这样一来，模型就有了它所需要的信息来决定下一步该做什么 548 | 549 | 138 550 | 00:07:52,680 --> 00:07:58,600 551 | 现在我们将设置并运行这个用户界面来显示订单机器人 552 | 553 | 139 554 | 00:07:58,600 --> 00:08:00,360 555 | 所以这里是上下文 556 | 557 | 140 558 | 00:08:00,360 --> 00:08:04,200 559 | 并且它包含了包含菜单的系统信息 560 | 561 | 141 562 | 00:08:04,560 --> 00:08:08,160 563 | 请注意，每次我们调用语言模型的时候 564 | 565 | 142 566 | 00:08:08,160 --> 00:08:13,400 567 | 我们都会使用相同的上下文，而且这个上下文是随着时间的推移而建立的 568 | 569 | 143 570 | 00:08:13,400 --> 00:08:19,040 571 | 然后让我们来执行这个 572 | 573 | 144 574 | 00:08:22,880 --> 00:08:25,760 575 | 我要说，你好 576 | 577 | 145 578 | 00:08:25,760 --> 00:08:29,440 579 | 我想订购一个比萨饼 580 | 581 | 146 582 | 00:08:33,080 --> 00:08:35,400 583 | 助理说，太好了 584 | 585 | 147 586 | 00:08:35,400 --> 00:08:36,880 587 | 你想点什么比萨饼？ 588 | 589 | 148 590 | 00:08:36,880 --> 00:08:40,520 591 | 我们有意大利香肠、奶酪和茄子比萨 592 | 593 | 149 594 | 00:08:40,520 --> 00:08:44,680 595 | 多少钱？ 596 | 597 | 150 598 | 00:08:53,080 --> 00:08:56,880 599 | 很好。好的，我们得到了价格 600 | 601 | 151 602 | 00:08:56,880 --> 00:09:12,000 603 | 我想我觉得是中号茄子比萨。所以你可以想象 604 | 605 | 152 606 | 00:09:12,000 --> 00:09:14,000 607 | 我们可以继续这个对话 608 | 609 | 153 610 | 00:09:14,000 --> 00:09:17,720 611 | 让我们看看我们在系统信息中放了什么 612 | 613 | 154 614 | 00:09:18,280 --> 00:09:20,400 615 | 所以你是orderbot 616 | 617 | 155 618 | 00:09:20,400 --> 00:09:23,280 619 | 为一家比萨店收集订单的自动化服务 620 | 621 | 156 622 | 00:09:23,280 --> 00:09:24,600 623 | 你首先要问候顾客 624 | 625 | 157 626 | 00:09:24,600 --> 00:09:27,800 627 | 然后收集订单，再问是自取还是配送 628 | 629 | 158 630 | 00:09:27,800 --> 00:09:29,480 631 | 你等待收集整个订单 632 | 633 | 159 634 | 00:09:29,480 --> 00:09:31,680 635 | 然后总结一下，最后进行检查 636 | 637 | 160 638 | 00:09:31,680 --> 00:09:33,800 639 | 如果顾客还想添加其他东西 640 | 641 | 161 642 | 00:09:33,800 --> 00:09:36,400 643 | 如果是送货，你可以要求提供地址 644 | 645 | 162 646 | 00:09:36,400 --> 00:09:38,040 647 | 最后，你拿到付款 648 | 649 | 163 650 | 00:09:38,040 --> 00:09:40,240 651 | 确保清晰描述所有选项、附加服务、额外费用 652 | 653 | 164 654 | 00:09:40,240 --> 00:09:43,120 655 | 和尺寸，以便从菜单上精确识别该项目 656 | 657 | 165 658 | 00:09:43,120 --> 00:09:44,920 659 | 你需要以简短的 660 | 661 | 166 662 | 00:09:44,920 --> 00:09:46,840 663 | 健谈的、友好的风格来回答 664 | 665 | 167 666 | 00:09:46,840 --> 00:09:50,880 667 | 菜单包括，然后在这里我们有菜单 668 | 669 | 168 670 | 00:09:51,120 --> 00:09:54,200 671 | 所以让我们回到我们的对话 672 | 673 | 169 674 | 00:09:54,200 --> 00:09:58,720 675 | 让我们看看助理是否一直在遵循指示 676 | 677 | 170 678 | 00:09:58,720 --> 00:10:01,720 679 | 好的，很好。助理问我们是否需要任何配料 680 | 681 | 171 682 | 00:10:01,720 --> 00:10:05,760 683 | 我们在系统信息中指定了这一点 684 | 685 | 172 686 | 00:10:05,760 --> 00:10:09,600 687 | 所以我想我们不需要额外的配料 688 | 689 | 173 690 | 00:10:15,600 --> 00:10:18,680 691 | 当然可以。还有什么我们想点的吗？ 692 | 693 | 174 694 | 00:10:18,680 --> 00:10:22,240 695 | 嗯。让我们来点水 696 | 697 | 175 698 | 00:10:22,240 --> 00:10:26,240 699 | 事实上，薯条 700 | 701 | 176 702 | 00:10:32,240 --> 00:10:34,240 703 | 小的还是大的？ 704 | 705 | 177 706 | 00:10:34,240 --> 00:10:36,240 707 | 这很好，因为我们在系统信息中 708 | 709 | 178 710 | 00:10:36,240 --> 00:10:41,240 711 | 要求助理在系统信息中说明额外的东西和配菜 712 | 713 | 179 714 | 00:10:46,240 --> 00:10:50,240 715 | 这样你就明白了，你可以随意试玩一下这个过程 716 | 717 | 180 718 | 00:10:50,240 --> 00:10:55,240 719 | 可以暂停视频，在左边的Notebook上运行这个 720 | 721 | 181 722 | 00:10:55,240 --> 00:10:59,240 723 | 现在我们可以要求该模型创建一个JSON摘要 724 | 725 | 182 726 | 00:10:59,240 --> 00:11:03,240 727 | 我们可以在对话的基础上发送至订单系统 728 | 729 | 183 730 | 00:11:03,240 --> 00:11:06,240 731 | 所以，我们现在要附加另一个系统消息 732 | 733 | 184 734 | 00:11:06,240 --> 00:11:08,240 735 | 这是一个指令 736 | 737 | 185 738 | 00:11:08,240 --> 00:11:11,240 739 | 我们在说创建一个关于之前食品订单的JSON摘要 740 | 741 | 186 742 | 00:11:11,240 --> 00:11:13,240 743 | 逐项列出每一项的价格 744 | 745 | 187 746 | 00:11:13,240 --> 00:11:18,240 747 | 字段应该是一个比萨饼，包括配菜 748 | 749 | 188 750 | 00:11:18,240 --> 00:11:21,240 751 | 两张配料表，三张饮料表 752 | 753 | 189 754 | 00:11:21,240 --> 00:11:23,240 755 | 和四份配菜清单 756 | 757 | 190 758 | 00:11:23,240 --> 00:11:25,240 759 | 最后是总价格 760 | 761 | 191 762 | 00:11:25,240 --> 00:11:29,240 763 | 你也可以在这里使用一个用户信息 764 | 765 | 192 766 | 00:11:29,240 --> 00:11:33,240 767 | 这不一定是一个系统消息 768 | 769 | 193 770 | 00:11:33,240 --> 00:11:36,240 771 | 所以让我们来执行这个 772 | 773 | 194 774 | 00:11:36,240 --> 00:11:39,240 775 | 注意在这种情况下，我们需要使用更低的temperature 776 | 777 | 195 778 | 00:11:39,240 --> 00:11:41,240 779 | 因为对于这些类型的任务 780 | 781 | 196 782 | 00:11:41,240 --> 00:11:43,240 783 | 我们希望输出是相当可预测的 784 | 785 | 197 786 | 00:11:43,240 --> 00:11:47,240 787 | 对于一个对话式的代理，你可能想使用更高的temperature 788 | 789 | 198 790 | 00:11:47,240 --> 00:11:50,240 791 | 然而，在这种情况下，我可能也会使用一个较低的temperature 792 | 793 | 199 794 | 00:11:50,240 --> 00:11:53,240 795 | 因为对于一个客户助理聊天机器人来说 796 | 797 | 200 798 | 00:11:53,240 --> 00:11:59,240 799 | 你可能想让输出更可预测一些 800 | 801 | 201 802 | 00:11:59,240 --> 00:12:02,240 803 | 因此，在这里我们有我们的订单摘要 804 | 805 | 202 806 | 00:12:02,240 --> 00:12:07,240 807 | 如果我们想的话，我们可以把这个提交给订单系统 808 | 809 | 203 810 | 00:12:07,240 --> 00:12:08,240 811 | 这就是我们需要的 812 | 813 | 204 814 | 00:12:08,240 --> 00:12:11,240 815 | 你已经建立了你自己的订单聊天机器人 816 | 817 | 205 818 | 00:12:11,240 --> 00:12:13,240 819 | 你可以自由地定制它 820 | 821 | 206 822 | 00:12:13,240 --> 00:12:16,240 823 | 试玩一下系统信息 824 | 825 | 207 826 | 00:12:16,240 --> 00:12:19,240 827 | 并改变聊天机器人的行为 828 | 829 | 208 830 | 00:12:19,240 --> 00:12:46,240 831 | 让它作为具有不同知识的不同角色行事 832 | 833 | --------------------------------------------------------------------------------