来自：口仆
进NLP群—>加入NLP交流群
本笔记是 deeplearning.ai 最近推出的短期课程《ChatGPT Prompt Engineering for Developers》的学习总结。
1 引言
总的来说，当前有两类大语言模型（LLM）：「基础 LLM」 和「指令微调 LLM」。
基础 LLM 基于大量文本数据训练而成，核心思想为预测一句话的下一个单词（即词语接龙）。基于语料的限制，有时会返回不符合预期的结果（如上图所示）。
指令微调 LLM 基于人类设定的指令（格式化的输入与输出）对基础 LLM 进行微调训练，使其能够返回更加有用且无害的结果。当前还会使用 RLHF（基于人类反馈的强化学习）来进一步提升模型效果，此处不做展开。
本课程旨在介绍如何基于 ChatGPT 官方 API 编写 Prompt（即指令），来驱动模型更好地完成各种各样的任务。
2 指南
总的来说，Prompt 的编写有两大原则：
原则一：编写明确而清晰的指令
原则二：给模型足够的时间思考
下面将结合具体的代码来对这两条原则进行详细说明。在这之前，首先给出本课程中所使用的模型调用函数的代码（略去了从环境变量获取 key 的部分）：
import openaiopenai.api_key = 'xxx'def get_completion(prompt, model="gpt-3.5-turbo"):messages = [{"role": "user", "content": prompt}]response = openai.ChatCompletion.create(model=model,messages=messages,temperature=0, # this is the degree of randomness of the model's output)return response.choices[0].message["content"]2.1 编写明确而清晰的指令
对于这一原则，课程中给出了四条具体的策略来实践该原则：
「策略 1：使用分隔符，避免 Prompt 入侵」
我们可以通过分隔符来标识仅需要 ChatGPT 进行阅读的文本，防止其被解析为指令（即 Prompt 入侵）。常用的分隔符包括：三个引号、三个反引号、三个破折号等，ChatGPT 对于这些分隔符的区分并不是很敏感（例如在 Prompt 中说明使用反引号分隔，但实际用的是引号，并不会影响模型输出）。
具体的代码示例如下（Prompt 不作翻译，可以自行替换为中文）：
text = f"""
内容省略（可添加任意内容）
"""
prompt = f"""
Summarize the text delimited by triple backticks \
into a single sentence.
```{text}```
"""response = get_completion(prompt)
print(response)「策略 2：要求模型提供结构化输出」
对于开发者来说，要求模型提供诸如 HTML 或 JSON 的格式化输出，有利于构建更加健壮的应用。具体的代码示例如下：
prompt = f"""
Generate a list of three made-up book titles along \
with their authors and genres.
Provide them in JSON format with the following keys:
book_id, title, author, genre.
"""response = get_completion(prompt)
print(response)「策略 3：让模型进行条件判断」
对于较复杂的 Prompt，模型生成结果的时间可能会较长，同时可能浪费大量的 token。为了规避不必要的 API 调用消耗，我们可以在 Prompt 中包含一定的条件判断逻辑，来帮助模型在不满足条件时提前终止运算，直接返回结果。具体的代码示例如下：
text_1 = f"""
如何把大象放进冰箱？\
首先打开冰箱的门，\
然后把大象放进去，\
最后关上冰箱门。
"""
prompt = f"""
You will be provided with text delimited by triple quotes.
If it contains a sequence of instructions, \
re-write those instructions in the following format:Step 1 - ...
Step 2 - …
…
Step N - …If the text does not contain a sequence of instructions, \
then simply write \"No steps provided.\"```{text_1}```
"""response = get_completion(prompt)
print("Completion for Text 1:")
print(response)「策略 4：少量（Few-shot）训练提示」
有时候，对于某些任务，我们需要为模型提供少量完成该任务的成功示例，来帮助模型更好地理解并完成该任务。具体的代码示例如下：
prompt = f"""
Your task is to answer in a consistent style.<child>: Teach me about patience.<grandparent>: The river that carves the deepest \
valley flows from a modest spring; the \
grandest symphony originates from a single note; \
the most intricate tapestry begins with a solitary thread.<child>: Teach me about love.
"""response = get_completion(prompt)
print(response)2.2 给模型足够的时间思考
这一原则的意义是：对于较为复杂的任务，如果不进行拆解，要求模型在短时间内通过少量词语进行回答，则可能导致结果的不准确。课程中给出了两条具体的策略来实践该原则：
「策略 1：指定完成任务所需的步骤」
该策略即通过 Prompt 拆解任务步骤降低复杂度，以帮助模型更好地完成任务。具体的代码示例如下（省略了 text 的内容，指定输出格式）：
prompt = f"""
Your task is to perform the following actions:
1 - Summarize the following text delimited by <> with 1 sentence.
2 - Translate the summary into Chinese.
3 - List each name in the Chinese summary.
4 - Output a json object that contains the following keys: chinese_summary, num_names.Use the following format:
Text: <text to summarize>
Summary: <summary>
Translation: <summary translation>
Names: <list of names in Chinese summary>
Output JSON: <json with summary and num_names>Text: <{text}>
"""response = get_completion(prompt)
print("\nCompletion for prompt 2:")
print(response)「策略 2：提示模型在匆忙得出结论之前思考自己的解决方案」
通过提示模型在回答之前思考自己的解决方案，有时可以得到更好的结果。课程中给出了一个解数学题的案例，如果没有提示模型首先尝试解题，则模型会判断学生的解法是正确的，但是如果提示模型首先自己推导解题过程再进行判断，则其会得出学生的解法是错误的结论。具体的代码如下（比较长）：
prompt = f"""
Your task is to determine if the student's solution \
is correct or not.
To solve the problem do the following:
- First, work out your own solution to the problem.
- Then compare your solution to the student's solution \
and evaluate if the student's solution is correct or not.
Don't decide if the student's solution is correct until
you have done the problem yourself.Use the following format:
Question:
'''
question here
'''
Student's solution:
'''
student's solution here
'''
Actual solution:
'''
steps to work out the solution and your solution here
'''
Is the student's solution the same as actual solution \
just calculated:
'''
yes or no
'''
Student grade:
'''
correct or incorrect
'''Question:
'''
I'm building a solar power installation and I need help \
working out the financials.
- Land costs $100 / square foot
- I can buy solar panels for $250 / square foot
- I negotiated a contract for maintenance that will cost \
me a flat $100k per year, and an additional $10 / square \
foot
What is the total cost for the first year of operations \
as a function of the number of square feet.
'''
Student's solution:
'''
Let x be the size of the installation in square feet.
Costs:
1. Land cost: 100x
2. Solar panel cost: 250x
3. Maintenance cost: 100,000 + 100x
Total cost: 100x + 250x + 100,000 + 100x = 450x + 100,000
'''
Actual solution:
"""response = get_completion(prompt)
print(response)该代码的输出结果为：
Let x be the size of the installation in square feet.Costs:
1. Land cost: 100x
2. Solar panel cost: 250x
3. Maintenance cost: 100,000 + 10xTotal cost: 100x + 250x + 100,000 + 10x = 360x + 100,000Is the student's solution the same as actual solution just calculated:
NoStudent grade:
IncorrectPS：课程中在多行文本中使用反斜杠来代替换行符，保证原始文本中没有换行，但是其在反斜杠添加了空格，会导致换行依旧存在，需要去除空格（实际对模型来说并没有影响，只是影响展示）
此外，本节还提到了模型的一种局限性：「幻觉」（Hallucination）。幻觉指 LLM 模型对于知识的边界性有时把握不足，可能会出现一本正经地胡说八道的情况。为了减少幻觉，我们可以提示模型首先收集相关信息，基于这些信息来回答问题。
3 迭代
编写 Prompt 是一个持续迭代的过程，通过对模型返回结果的分析，不断地修改 Prompt，我们可以最终得到较为满意的输出。同时，对 Prompt 的修改需要遵循上一章节提到的原则。
4 摘要
我们可以使用 ChatGPT 来辅助生成对长文本的总结，以帮助我们在短时间内获取更多的信息。具体来说，通过模型执行摘要任务有以下两个注意点：
4.1 细化需求
通过 Prompt 来细化摘要的具体目的与关注点，可以获得更加精准的摘要。具体的代码示例如下：
prompt = f"""
Your task is to generate a short summary of a product \
review from an ecommerce site to give feedback to the \
Shipping deparmtment. Summarize the review below in Chinese, delimited by triple
backticks, in at most 30 words, and focusing on any aspects \
that mention shipping and delivery of the product. Review: '''{prod_review}'''
"""response = get_completion(prompt)
print(response)4.2 关键字差异
如果在 Prompt 中使用关键字「总结」（summarize），虽然模型会基于 Prompt 返回对应的总结，但其中通常会包含一些其他的信息；而如果使用关键字「提取」（extract），则模型会专注于提取在 prompt 中所提示的范围，返回更加精准的摘要。具体的代码示例如下：
prompt = f"""
Your task is to extract relevant information from \
a product review from an ecommerce site to give \
feedback to the Shipping department. The extracted result \
should be translated into Chinese.From the review below delimited by triple quotes \
extract the information relevant to shipping and \
delivery. Limit to 30 words. Review: '''{prod_review}'''
"""response = get_completion(prompt)
print(response)5 推理
ChatGPT 非常适合用来对文本进行推理，对于传统 NLP 方法来说需要大量步骤（例如数据收集、打标签、训练、测试等）才能完成的任务，ChatGPT 可以通过 Prompt 轻松完成。本节将列举两种主要的推理任务：「情感分析」与「信息提取」。
5.1 情感分析
我们可以通过多种方式来指导 ChatGPT 分析一段文字中的情感，例如：
「识别积极/消极」
prompt = f"""
What is the sentiment of the following product review,
which is delimited with triple backticks?Give your answer as a single word, either "positive" \
or "negative".Review text: '''{lamp_review}'''
"""response = get_completion(prompt)
print(response)「识别多种情感」
prompt = f"""
Identify a list of emotions that the writer of the \
following review is expressing. Include no more than \
five items in the list. Format your answer as a list of \
lower-case words separated by commas.Review text: '''{lamp_review}'''
"""response = get_completion(prompt)
print(response)「识别特定情感」
prompt = f"""
Is the writer of the following review expressing anger?\
The review is delimited with triple backticks. \
Give your answer as either yes or no.Review text: '''{lamp_review}'''
"""
response = get_completion(prompt)
print(response)5.2 信息提取
我们可以通过不同的 Prompt 指导 ChatGPT 提取不同维度的信息，例如：
「提取标签（实体）信息」
prompt = f"""
Identify the following items from the review text:
- Item purchased by reviewer
- Company that made the itemThe review is delimited with triple backticks. \
Format your response as a JSON object with \
"Item" and "Brand" as the keys.
If the information isn't present, use "unknown" \
as the value.
Make your response as short as possible.Review text: '''{lamp_review}'''
"""response = get_completion(prompt)
print(response)「提取主题信息」
对于主题提取，ChatGPT 可以直接提取多个主题：
prompt = f"""
Determine five topics that are being discussed in the \
following text, which is delimited by triple backticks.Make each item one or two words long. Format your response as a list of items separated by commas.Text sample: '''{story}'''
"""response = get_completion(prompt)
print(response)也可以给定主题列表，判断输入文本中包含了哪些特定主题（该方式也被称为 Zero Shot 训练）：
topic_list = ["nasa", "local government", "engineering", "employee satisfaction", "federal government"
]prompt = f"""
Determine whether each item in the following list of \
topics is a topic in the text below, which
is delimited with triple backticks.Give your answer as list with 0 or 1 for each topic.\List of topics: {", ".join(topic_list)}Text sample: '''{story}'''
"""response = get_completion(prompt)
print(response)6 转换
本节将介绍如何使用 ChatGPT 来进行文本转换任务，例如翻译、拼写与语法检查、语气调整、格式转换等，主要以代码示例为主。
6.1 翻译
ChatGPT 基于多种语言进行训练，具有较强的翻译能力，例如：
「翻译多种语言」
prompt = f"""
Translate the following text to French and Spanish
and English pirate: \
'''I want to order a basketball'''
"""response = get_completion(prompt)
print(response)「识别当前语言」
prompt = f"""
Tell me which language this is:
'''Combien coûte le lampadaire?'''
"""response = get_completion(prompt)
print(response)6.2 语气调整
基于不同的读者（观众），输出的文字可能有较大差异。ChatGPT 可以基于 Prompt 产生不同的语调，例如：
prompt = f"""
Translate the following from slang to a business letter:
'Dude, This is Joe, check out this spec on this standing lamp.'
"""response = get_completion(prompt)
print(response)6.3 格式转换
ChatGPT 支持多种格式的转换，Prompt 应该清楚地描述输入与输出格式，例如：
data_json = { "resturant employees" :[ {"name":"Shyam", "email":"shyamjaiswal@gmail.com"},{"name":"Bob", "email":"bob32@gmail.com"},{"name":"Jai", "email":"jai87@gmail.com"}
]}prompt = f"""
Translate the following python dictionary from JSON to an HTML \
table with column headers and title: {data_json}
"""response = get_completion(prompt)
print(response)6.4 文字检查
ChatGPT 还可以用来检查拼写和语法，帮助你改写文字（写文章必备），例如：
「普通校对与改写」
text = f"""
Got this for my daughter for her birthday cuz she keeps taking \
mine from my room. Yes, adults also like pandas too. She takes \
it everywhere with her, and it's super soft and cute. One of the \
ears is a bit lower than the other, and I don't think that was \
designed to be asymmetrical. It's a bit small for what I paid for it \
though. I think there might be other options that are bigger for \
the same price. It arrived a day earlier than expected, so I got \
to play with it myself before I gave it to my daughter.
"""prompt = f"proofread and correct this review: ```{text}```"
response = get_completion(prompt)
print(response)课程中还使用了 redlines 包来查看改写前后的差异：
from redlines import Redlinesdiff = Redlines(text,response)
display(Markdown(diff.output_markdown))
「基于特定风格的校对与改写」
prompt = f"""
proofread and correct this review. Make it more compelling.
Ensure it follows APA style guide and targets an advanced reader.
Output in markdown format.
Text: ```{text}```
"""response = get_completion(prompt)
display(Markdown(response))7 扩展
ChatGPT 可以用来对文字进行扩展（将短文本变成长文本），本节将给出一个使用 ChatGPT 自动回复 email 的示例，并介绍对扩展来说很重要的一个 API 参数：「温度」（Temperature）
7.1 扩展示例
下面的例子给出了基于用户的评价生成自动回复 email 的 Prompt：
prompt = f"""
You are a customer service AI assistant.
Your task is to send an email reply in Chinese to a valued customer.
Given the customer email delimited by ```, \
Generate a reply to thank the customer for their review.
If the sentiment is positive or neutral, thank them for \
their review.
If the sentiment is negative, apologize and suggest that \
they can reach out to customer service.
Make sure to use specific details from the review.
Write in a concise and professional tone.
Sign the email as `AI customer agent`.
Customer review: '''{review}'''
Review sentiment: {sentiment}
"""response = get_completion(prompt)
print(response)实际上情感可以由 ChatGPT 自动提取，无需指定。此外，为了防止不好的用途，课程中建议明确表明上述内容是由 AI 生成的。
7.2 温度参数
温度参数可以用来控制模型输出的随机性，具体规则如下：
温度越低，模型的输出越稳定（更保守）
温度越高，模型的输出越随机（更具有创造力）
在不同的场景下，通过对温度参数的控制，可以得到更加合适的输出（温度参数通常设置在 0-2 之间）。
8 聊天机器人
本节将介绍如何使用 ChatGPT 构建一个聊天机器人，其关键在于如何为模型提供「上下文信息」。对模型来说，每一轮对话都是独立的，我们需要在接口中传入之前对话的历史信息。对于接口来说，总共存在三种角色，分别是：
System 角色：用于设定模型的身份，约束模型行为，放置于一轮对话的最开头
User 角色：当前对话的用户，提供输入信息
Assistant 角色：生成回复的模型
历史信息需要明确地表明每一条消息所对应的角色，如下图所示：
为了实现历史信息的传入，需要修改之前的模型调用函数：
def get_completion_from_messages(messages, model="gpt-3.5-turbo", temperature=0):response = openai.ChatCompletion.create(model=model,messages=messages,temperature=temperature, # this is the degree of randomness of the model's output)
# print(str(response.choices[0].message))return response.choices[0].message["content"]基于上述函数，可以手动实现历史消息的传入：
messages = [
{'role':'system', 'content':'You are friendly chatbot.'},
{'role':'user', 'content':'Hi, my name is Isa'},
{'role':'assistant', 'content': "Hi Isa! It's nice to meet you. \
Is there anything I can help you with today?"},
{'role':'user', 'content':'Yes, you can remind me, What is my name?'} ]
response = get_completion_from_messages(messages, temperature=1)
print(response)为了实现一个聊天机器人，我们需要支持对于对话历史消息的自动收集，课程中基于 panel 包实现了一个点菜机器人，通过精心设置的 System Prompt 实现了自动化点菜与价格计算，具体的代码此处不作展开。
这里给出一个更加通用的聊天机器人 demo，基于 Streamlit 实现，可以自由设置 System Prompt，同时对于对话历史消息进行了自动收集与上传：
import streamlit as st
import openaist.markdown('# ChatGPT demo')def init_session():if 'round' not in st.session_state:print("重置轮数")st.session_state['round'] = 1if 'question' not in st.session_state:print("重置问题记录")st.session_state['question'] = []if 'answer' not in st.session_state:print("重置回答记录")st.session_state['answer'] = []init_session()
code_input = st.text_input('请输入使用密钥')
system_input = st.text_input('请设定 ChatGPT 角色（用于生成符合角色的回答）', "你是一个有用的AI助手")
latest_input = st.text_area('请输入内容')openai.api_key = code_inputif st.button('发送'):st.session_state['question'].append(latest_input)message_list = []dic_q_item = {"role": "system", "content": system_input}message_list.append(dic_q_item)for n in range(st.session_state['round']):# 第 n 轮，有 n 个问题，n - 1 个答案待组装if n - 1 >= 0:dic_a_item = {"role": "assistant", "content": st.session_state['answer'][n - 1]}message_list.append(dic_a_item)dic_q_item = {"role": "user", "content": st.session_state['question'][n]}message_list.append(dic_q_item)completion = openai.ChatCompletion.create(model="gpt-3.5-turbo",max_tokens=1000,user="wzy",messages=message_list,temperature=0)answer = completion.choices[0].message['content']st.session_state['answer'].append(answer)col1, col2 = st.columns(2)for i in range(st.session_state['round']):with col1:st.text('User')st.write(' ' + st.session_state['question'][i])st.text(' ')st.write(' ')with col2:st.text(' ')st.write(' ')st.text('ChatGPT')st.write(' ' + st.session_state['answer'][i])st.session_state['round'] += 1if st.button('清空对话记录'):st.session_state.clear()9 总结
以下为本课程核心知识点的思维导图总结：
进NLP群—>加入NLP交流群
