ChatGPT API
簡單的說， model gpt-3.5-turbo 就是 對話體。新增的 API ChatCompletion，提供了新的訊息格式（其他參數還是可以像之前那樣設定）。
你可以使用這個 API，創造出類似 ChatGPT 的 ChatBot：一問一答，而且可以記住之前的對話（連續性對話，我認為勝過 google 搜尋的優點，感興趣的朋友可以參考這篇： Google search vs. ChatGPT ）。
ChatGPT 模型針對 對話 進行了最佳化，gpt-3.5-turbo 的效率與 Instruct Davinci 相當。
ChatGPT API 無法進行 fine-tuned（微調，不過未來可能會改變），呼叫的方式也不同。我們看官網 OpenAI API 的範例。
範例程式
# Note: you need to be using OpenAI Python v0.27.0 for the code below to work
import openai
openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Who won the world series in 2020?"},
{"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."},
{"role": "user", "content": "Where was it played?"}
]
)
範例中的 messages 有四部分：
| role | content |
|---|---|
| system | You are a helpful assistant. |
| user | Who won the World Series in 2020? |
| assistant | The Los Angeles Dodgers won the World Series in 2020. |
| user | Where was it played? |
首先設定 system，讓 ChatCompletion API 知道要以什麼身份回答問題。就像大家在許多文章看到的，ChatGPT 的 prompt 撰寫技巧：你現在是一個…。
接著第二段第三段分別是使用者的問題，和 ChatGPT 的回答。
值得留意的是第四段使用者的新問題：Where was it played?
當你使用 ChatCompletion API，把四段訊息串在一起時，OpenAI 會知道這裡的 Where 指的是前一問句的相關問題。
提醒：只要不超過文字上限，
messages可以一直串接下去，只有第一段是system，接下來就是user和assistant輪流對話。
補充資料
- 需升級到 0.27.0 以上的版本，如果沒有的話，以下方指令升級：
pip install openai --upgrade
價格
與之前 Models 的價格比較
2023年4月22日資料
| Model | Training | Usage | ||
|---|---|---|---|---|
| gpt-3.5-turbo | $0.0300／1K tokens | $0.0020／1K tokens | ||
| Davinci | $0.0300／1K tokens | $0.1200／1K tokens | ||
| Ada | $0.0004／1K tokens | $0.0016／1K tokens | ||
| Babbage | $0.0006／1K tokens | $0.0024／1K tokens | ||
| Curie | $0.0030／1K tokens | $0.0120／1K tokens |
為什麼 gpt-3.5-turbo 特別便宜呢？我推測是因為 messages 需要一直重覆之前的對話串。
RHLF
RHLF: Reinforcement Learning from Human Feedback
人類回饋強化學習
WIKI
論文 Paper
Learning to summarize with human feedback
ChatGPT Model Comparison
GPT-3.5
GPT-3.5 models can understand and generate natural language or code. Our most capable and cost effective model in the GPT-3.5 family is gpt-3.5-turbo which has been optimized for chat but works well for traditional completions tasks as well.
| LATEST MODEL | DESCRIPTION | MAX TOKENS | TRAINING DATA |
|---|---|---|---|
| gpt-3.5-turbo | Most capable GPT-3.5 model and optimized for chat at 1/10th the cost of text-davinci-003. Will be updated with our latest model iteration. |
4,096 tokens | Up to Sep 2021 |
| gpt-3.5-turbo-0301 | Snapshot of gpt-3.5-turbo from March 1st 2023. Unlike gpt-3.5-turbo, this model will not receive updates, and will be deprecated 3 months after a new version is released. |
4,096 tokens | Up to Sep 2021 |
| text-davinci-003 | Can do any language task with better quality, longer output, and consistent instruction-following than the curie, babbage, or ada models. Also supports inserting completions within text. | 4,097 tokens | Up to Jun 2021 |
| text-davinci-002 | Similar capabilities to text-davinci-003 but trained with supervised fine-tuning instead of reinforcement learning |
4,097 tokens | Up to Jun 2021 |
| code-davinci-002 | Optimized for code-completion tasks | 8,001 tokens | Up to Jun 2021 |
We recommend using gpt-3.5-turbo over the other GPT-3.5 models because of its lower cost.
資料來源：
https://platform.openai.com/docs/models/gpt-3-5
專案實作
以之前的範例 History Tutor ChatBot Remix 修改，改用 ChatCompletion 實作
本案例幾個改進原專案的點：
-
呼叫的 OpenAI API 改為
ChatCompletion -
以一個 while 迴圈處理連續對話
-
計算 token 以免看到帳單昏倒
-
在 prompt 中加入相關敘述防呆，避免 hallucination 誤判（細節請參考前一節課程 文字嵌入（Text Embedding））
import openai
import os
openai.api_key = os.getenv("OPENAI_API_KEY")
class CreateBot:
def __init__(self, system_prompt):
'''
system_prompt: [str] Describes context for Chat Assistant
'''
self.system = system_prompt
self.messages = [{"role": "system", "content": system_prompt}]
def chat(self):
'''
Tracks dialogue history and takes in user input
'''
print('To end conversation, type END')
question = ''
while question != 'END':
# Get User Question
question = input("")
# Add to messages/dialogue history
self.messages.append({'role':'user','content':question})
#Send to ChatGPT and get response
response = openai.ChatCompletion.create(
model="gpt-3.5-turbo",
messages=self.messages)
# Get content of assistant reply
content = response['choices'][0]['message']['content']
print('\n')
print(content)
print('\n')
# Add assistant reply for dialogue history
self.messages.append({'role':'assistant','content':content})
history_tutor = CreateBot(system_prompt="You are an expert in US History")
history_tutor.chat()
Token Length Check
full_content = ''
for item in history_tutor.messages:
full_content += item['content']
def num_tokens_from_string(string, encoding_name):
"""Returns the number of tokens in a text string."""
encoding = tiktoken.get_encoding(encoding_name)
num_tokens = len(encoding.encode(string))
return num_tokens
num_tokens_from_string(full_content,encoding_name='cl100k_base')
Out（輸出）
485
# Don't pop index 0, because that is the system role!
history_tutor.messages.pop(1)
Out（輸出）
{‘role’: ‘user’, ‘content’: ‘Who was the first president of the US?’}
