OpenAI GPT LLMs 高级提示词工程方法汇总

admin管理员组

文章数量:1584248

原文地址：An Introduction to Prompt Engineering for OpenAI GPT LLMs

Github：Prompt-Engineering-Intro

2023 年 3 月 2 日

提示工程指南 | Prompt Engineering Guide

Naive 提示词：带有提示的情感分类器

prompt = '''
Decide whether a Tweet's sentiment is positive, neutral, or negative.

Tweet: "I loved the new Batman movie!"
Sentiment:
'''

Advanced 提示词

为了构建一个问答机器人，我们可以提供文档或废弃的网站作为上下文信息。然后，我们可以向模型提出问题，以引导它通过必要的中间步骤来构建答案。此外，我们还可以要求模型生成代码来执行特定任务，比如集成外部工具（例如 Python 代码执行器、API 调用、数据库查询等）。这些代码的输出可以作为历史记录，用于提出更深入的问题，并作为新的提示来进一步引导模型。

• Few Shot Prompting
• Chain of Thoughts Prompting
• Self Consistency Prompting
• ReAct — Reason and Act prompting
• PAL — Program aided Language Models
• MRKL Systems — Modular Reasoning, Knowledge and Language
• Self-Ask with Search

Few Shot

Few-Shot Prompting是一种提示策略，它向语言模型提供少量的示例作为输入的一部分，以指导模型生成所需的输出。在这种提示中，通常包括三个子部分：

• 1. 前缀（Instruction）：这部分是对模型的指示，告诉模型需要完成的任务。例如，“给出每个输入的反义词”。
• 2. 示例（Examples）：这部分包含了一些完成任务的具体示例，这些示例展示了任务的预期输出。例如，一个单词和它的反义词的列表。
• 3. 后缀（User Input）：这部分是用户的实际查询，即模型需要为其生成输出的输入。在这个例子中，就是用户希望得到反义词的那个单词。

Few-Shot之所以得名，是因为它只使用了少量的示例来“提示”模型，而不是对模型进行完整的重新训练。在这种方法中，模型是静态的，不会因为新的任务而更新其参数。相反，我们通过构建包含示例的提示来引导模型产生期望的结果。这种方法利用了大型语言模型的能力，即它们可以从少量的示例中快速学习并泛化到新的情况。

examples = [
    {"word": "happy", "antonym": "sad"},
    {"word": "tall", "antonym": "short"},
    {"word": "sunny", "antonym": "cloudy"}
]

example_formatter_template = """
Word: {word}
Antonym: {antonym}\n
"""

example_prompt = PromptTemplate(
    input_variables=["word", "antonym"],
    template=example_formatter_template,
)

few_shot_prompt = FewShotPromptTemplate(
    examples=examples,
    example_prompt=example_prompt,
    suffix="Add three other examples.",
    input_variables=[],
)

llm = OpenAI()
chain = LLMChain(llm=llm, prompt=few_shot_prompt)
print("### Model Output ###")
print(chain.predict())

######################################################

### Model Output ###

Word: fast
Antonym: slow

Word: generous
Antonym: stingy

Word: strong
Antonym: weak

Question: When was the last FIFA World Cup held?
Answer:
Is follow-up needed? Yes, follow-up questions are needed.
Follow-up: Which country hosted the last FIFA World Cup? Intermediate Answer: The last FIFA World Cup was hosted by Qatar.
Follow-up: Who won the last FIFA World Cup? Intermediate Answer: The last FIFA World Cup was won by Argentina.
Follow-up: When did the last FIFA World Cup take place? Intermediate Answer: The last FIFA World Cup took place in 2022.
So the final answer is: The last FIFA World Cup was held in Qatar in 2022, and Argentina emerged as the winner.

Question: Who is considered the greatest basketball player of all time?
Answer:
Are follow-up questions required? Yes, follow-up questions are needed.
Follow-up: How many NBA championships has the greatest basketball player won? Intermediate
Answer: The greatest basketball player, Michael Jordan, won six NBA championships.
Follow-up: Which team did Michael Jordan spend the majority of his career with? Intermediate
Answer: Michael Jordan spent the majority of his career with the Chicago Bulls.
So the final answer is: Michael Jordan is considered the greatest basketball player of all time, winning six NBA championships with the Chicago Bulls.

Question : Who won the 2023 cricket world cup?
Answer:

Zero Shot

在提示中，添加 ‘think step by step’，‘break this task into simpler subtask and solve’等短语作为原始提示旁边的后缀/前缀。这样的提示可以是这样的：

 “Think step by step and explain the ste

本文标签： 提示高级方法工程openAI