【李宏毅-生成式AI】Spring 2024, HW3：以API快速搭建自己的应用

实验要求使用Google Gemini API或者ChatGPT (OpenAI API)搭建自己的应用。

1. 准备工作

学会使用Google Colab写代码，并使用Gradio来托管App。

详见课程Slides

2. 获取API key

可选的API有：

• Google Gemini API：免费；
• ChatGPT (OpenAI API)：有5美元的免费额度；

PS：这两个在中国大陆都不可用，请自行科学上网。

“获取API Key”详见课程Slides。

3. Task

目标：了解如何通过调用API和输入prompts来构建自己的Language Model应用。

Task：

• Summarization (3 points)

• Role Playing (3 points)

• Customized Task (4 points)

填入API key

本文使用Gemini API进行示范。

# Import packages
import google.generativeai as genai
from typing import List, Tuple
import gradio as gr
import json

# Set up Gemini API key
## TODO: Fill in your Gemini API in the ""
GOOGLE_API_KEY=""
genai.configure(api_key=GOOGLE_API_KEY)
model = genai.GenerativeModel('gemini-pro')

# Check if you have set your Gemini API successfully
# You should see "Set Gemini API sucessfully!!" if nothing goes wrong.
try:
    model.generate_content(
      "test",
    )
    print("Set Gemini API sucessfully!!")
except:
    print("There seems to be something wrong with your Gemini API. Please follow our demonstration in the slide to get a correct one.")

将API key填入GOOGLE_API_KEY即可。

Task 1: Summarization

设计一个prompt，使语言模型能够对文章进行总结。

# function to call the model to generate
def interact_summarization(prompt: str, article: str, temp = 1.0) -> List[Tuple[str, str]]:
    '''
      * Arguments

        - prompt: the prompt that we use in this section

        - article: the article to be summarized

        - temp: the temperature parameter of this model. Temperature is used to control the output of the chatbot.
                The higher the temperature is, the more creative response you will get.
    '''
    input = f"{prompt}\n{article}"
    response = model.generate_content(
      input,
      generation_config=genai.types.GenerationConfig(temperature=temp),
      safety_settings=[
          {"category": "HARM_CATEGORY_HARASSMENT","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_HATE_SPEECH","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_DANGEROUS_CONTENT","threshold": "BLOCK_NONE",},
          ]
    )

    return [(input, response.text)]

将{prompt}和{article}拼接为input，传给Gemini API model.generate_content，返回的response即为语言模型生成的结果。

Task 2: Role Playing

设计一个机器人服务，可以和LM一起玩角色扮演游戏。要求与LM进行多轮对话。

# function to call the model to generate
def interact_roleplay(chatbot: List[Tuple[str, str]], user_input: str, temp=1.0) -> List[Tuple[str, str]]:
    '''
    * Arguments

      - user_input: the user input of each round of conversation

      - temp: the temperature parameter of this model. Temperature is used to control the output of the chatbot.
              The higher the temperature is, the more creative response you will get.

    '''
    try:
        messages = []
        for input_text, response_text in chatbot:
            messages.append({'role': 'user', 'parts': [input_text]})
            messages.append({'role': 'model', 'parts': [response_text]})

        messages.append({'role': 'user', 'parts': [user_input]})

        response = model.generate_content(
          messages,
          generation_config=genai.types.GenerationConfig(temperature=temp),
          safety_settings=[
          {"category": "HARM_CATEGORY_HARASSMENT","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_HATE_SPEECH","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_DANGEROUS_CONTENT","threshold": "BLOCK_NONE",},
          ]
        )

        chatbot.append((user_input, response.text))


    except Exception as e:
        print(f"Error occurred: {e}")
        chatbot.append((user_input, f"Sorry, an error occurred: {e}"))
    return chatbot

第一轮：input_text对应输入的prompt，调用model.generate_content返回生成的回答response.text；

第二轮：input_text对应输入的prompt，response_text对应第一轮的response.text。新的prompt对应

user_input。

调用model.generate_content之后，user_input加上 response.text组成完整的dialogue。

Task 3: Customize Task

创建一个自定义的服务机器人。例如，例如：一个可以解决简单数学问题的机器人、始终输出用户输入单词反义词的机器人等。

# function to call the model to generate
def interact_customize(chatbot: List[Tuple[str, str]], prompt: str ,user_input: str, temp = 1.0) -> List[Tuple[str, str]]:
    '''
    * Arguments

      - chatbot: the model itself, the conversation is stored in list of tuples

      - prompt: the prompt for your desginated task

      - user_input: the user input of each round of conversation

      - temp: the temperature parameter of this model. Temperature is used to control the output of the chatbot.
              The higher the temperature is, the more creative response you will get.

    '''
    try:
        messages = []

        for input_text, response_text in chatbot:
            messages.append({'role': 'user', 'parts': [input_text]})
            messages.append({'role': 'model', 'parts': [response_text]})

        messages.append({'role': 'user', 'parts': [prompt+ "\n" + user_input]})

        response = model.generate_content(
          messages,
          generation_config=genai.types.GenerationConfig(temperature=temp),
          safety_settings=[
          {"category": "HARM_CATEGORY_HARASSMENT","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_HATE_SPEECH","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_SEXUALLY_EXPLICIT","threshold": "BLOCK_NONE",},
          {"category": "HARM_CATEGORY_DANGEROUS_CONTENT","threshold": "BLOCK_NONE",},
          ]
        )

        chatbot.append((user_input, response.text))

    except Exception as e:
        print(f"Error occurred: {e}")
        chatbot.append((user_input, f"Sorry, an error occurred: {e}"))
    return chatbot

def interact_customize()传入用户prompt，使语言模型成为“服务机器人”。后续的用户输入对应input_text，模型的返回为response_text。