프라이빗 GPT 모델 만들기

#프라이빗 #GPT #모델 #만들기 #OpenAI #LLM #LangChain #LIamaIndex

 

• 자동화를 위한 일관된 답 얻기와 데이터 처리 방법
• 인식 모델과 생성 모델의 차이와 딥러닝 모델의 한계
• 자동화를 위한 컴플릿리과 체의 차이와 펑션 정의, 임베딩의 의미

 

- 임베딩의 개념과 활용

- 모델 평가 방법과 객관적 평가 방법

- 버전 관리와 업그레이드의 중요성

- 인공지능 트랜스포머와 메모리 세팅의 차이

- 챗봇 개발과 위키피디아 임베딩

- 데이터 중복 제거와 API 사용의 필요성

- 라마 인덱스와 데이터 처리 기능 제공

- 오픈 ai와 코드 대문자화의 필요성

- sa를 통한 질문과 유용성

- 인덱스와 랭크를 이용한 상장회사 공시 보고서 검색

- 인공지능 모델의 전문성과 범형성

- 생성형 ai를 활용한 새로운 방법의 전달

 

CAPABILITIES

- Text generation

- Function calling

- Embeddings

- Fine-tuning

- Image generation

- Vision

- Text-to-speech

- Speech-to-text

- Moderation

 

https://gptforwork.com/tools/openai-api-and-other-llm-apis-response-time-tracker

OpenAI API and other LLM APIs response time tracker

The 3 charts below track the response times of the main large language model APIs: OpenAI and Azure OpenAI (GPT-4, GPT-3.5, GPT-3), Anthropic Claude and Google PaLM.

gptforwork.com

 

https://status.openai.com/uptime

OpenAI Status - Uptime History

Incidents Uptime ← Current Status Powered by Atlassian Statuspage

status.openai.com

 

 

 

 

 

 

 

 

인식모델 -> 생성모델

 

인식모델

https://cloud.google.com/vision

Vision AI | Google Cloud

Derive insights from images with AutoML Vision, or use pre-trained Vision API models or create computer vision applications with Vertex AI Vision

cloud.google.com

인식모델은 케익을 가방으로 인식

 

생성모델 vs 인식모델

 

생성모델

 

https://bard.google.com/chat/

 

스포츠 중계 방송 자동화 등

 

 

 

 

 

 

 

 

 

 

 

Embeddings

 

https://atlas.nomic.ai/map/3e58c76e-6cf4-4c3e-9184-38b174cbf3e8/a46f4c16-8511-4fde-9bf9-dcd1f8c72c0a

groovy-vein | Atlas

Structure unstructured datasets of text, images, embeddings, audio and video.

atlas.nomic.ai

 

 

 

 

 

 

 

Function calling

 

 

 

 

- Function calling

- Embeddings

- Fine-tuning

 

 

 

 

 

 

- Fine-tuning

 

 

RAG applications can use the LlamaIndex query interface for accessing that data and powering LLMs

 

LlamaIndex vs LangChain

 

 

Fine-tuning improves on few-shot learning by training on many more examples than can fit in the prompt

 

https://www.youtube.com/watch?v=WsMB0hOEosI

 

728x90

 

 

 

 

 

%%capture
!pip install openai==0.28
!pip install datasets

 

!pip show openai | grep Version

 

import os
import getpass
import openai
openai.api_key = getpass.getpass(prompt = 'OpenAI API')

 

dataset = {
    "train": [
      {
          "instruction": " ",
          "output": " "
      },
      {
          "instruction": " ",
          "output": " "
      } 
    ]
}

 

import json

list_message = []
num_data = len(dataset["train"])  # 데이터셋의 길이를 사용

for i in range(num_data):
    instruction = dataset["train"][i]["instruction"]
    output = dataset["train"][i]["output"]
    print("질문:", instruction)
    print("답변:", output)
    message = [
        {"role": "user", "content": instruction},
        {"role": "assistant", "content": output},
    ]
    list_message.append(message)

with open("output1.jsonl", "w") as file:
    for messages in list_message:
        json_line = json.dumps({"messages": messages})
        file.write(json_line + '\n')

 

 

upload_file = openai.File.create(
  file=open("output1.jsonl", "rb"),
  purpose='fine-tune'
)

 

import openai

openai.api_key = 'sk-'  # Replace with your actual API key

try:
    upload_file = openai.File.retrieve("file-O0eEwlZRoFvDANDKqBatdExq")
    print(upload_file)
except openai.error.InvalidRequestError as e:
    print("Error occurred:", e)

 

start_train = openai.FineTuningJob.create(training_file=upload_file["id"],model="gpt-3.5-turbo")

 

# Fine-tuning 작업 상태 확인
status = openai.FineTuningJob.list(limit=10)
status["data"][0]

 

# 모델 이름 얻기
fine_tuned_model = status["data"][1]["fine_tuned_model"]
print(fine_tuned_model)

 

completion = openai.ChatCompletion.create(
  model=status["data"][1]["fine_tuned_model"],
  messages=[
    {"role": "user", "content": ""}
  ]
)

 

print(completion.choices[0].message["content"])

 

completion = openai.ChatCompletion.create(
  model=status["data"][1]["fine_tuned_model"],
  messages=[
    {"role": "user", "content": "Python의 @에 대한 설명"}
  ]
)

 

 

completion = openai.ChatCompletion.create(
  model=status["data"][1]["fine_tuned_model"],
  messages=[
    {"role": "user", "content": "라면 끓일때 가장 중요한 것은?"}
  ]
)

 

print(completion.choices[0].message["content"])

 

 

- Fine-tuning 파인튜닝

 

 

 

 

 

- Pretraining 프리트레이닝 모델

 

LangChain

파이썬 문서: https://python.langchain.com/en/latest/index.html

개념 문서: https://docs.langchain.com/docs/

소스: https://github.com/hwchase17/langchain

 

#@title 기본 패키지(openai, langchain) 설치
!pip install openai==0.28
!pip install langchain

#@title 기타 패키지 설치 (구글검색, 위키피디아, VectorStore, HuggingFace Embedding)
!pip install google-search-results
!pip install wikipedia
!pip install faiss-cpu # 오픈소스 벡터DB (Facebook, MIT license)
!pip install sentence_transformers # HuggingFace Embedding 사용 위해서 필요
!pip install tiktoken # Summarization 할때 필요

#@title 0. API 키 설정
import os
#@markdown https://platform.openai.com/account/api-keys
OPENAI_API_KEY = "sk-" #@param {type:"string"}

#@markdown https://huggingface.co/settings/tokens
#@markdown HuggingFace에서 모델 다운로드나 클라우드 모델 사용하기 위해서 필요 (무료)
HUGGINGFACEHUB_API_TOKEN = "hf_" #@param {type:"string"}

#@markdown https://serpapi.com/manage-api-key
#@markdown 구글 검색하기 위해서 필요 (월 100회 무료)
SERPAPI_API_KEY = "" #@param {type:"string"}

os.environ["OPENAI_API_KEY"] = OPENAI_API_KEY
os.environ["HUGGINGFACEHUB_API_TOKEN"] = HUGGINGFACEHUB_API_TOKEN
os.environ["SERPAPI_API_KEY"] = SERPAPI_API_KEY

#@title 1. OpenAI LLM (text-davinci-003)
from langchain.llms import OpenAI

#@title 1. OpenAI LLM (text-davinci-003)
from langchain.llms import OpenAI
llm = OpenAI(model_name='text-davinci-003', temperature=0.9)

langchain.llms.openai.OpenAI

llm('미국의 빌보드차트에 오른 한국 가수 이름을 알려 주세요.')

llm('2020 올림픽은 어디에서 개최 되었나요?')

#@title 2. ChatOpenAI LLM (gpt-3.5-turbo)

from langchain.chat_models import ChatOpenAI
from langchain.schema import (
    AIMessage,
    HumanMessage,
    SystemMessage
)

chat = ChatOpenAI(model_name='gpt-3.5-turbo', temperature=0.9)
sys = SystemMessage(content="당신은 스포츠 전문 AI입니다.")
msg = HumanMessage(content='2020 하계 올림픽에서 대한민국 금메달 획득 갯수는?')

aimsg = chat([sys, msg])
aimsg.content

#@title 3. Prompt Template & chain
from langchain.prompts import PromptTemplate

prompt = PromptTemplate(
    input_variables=["상품"],
    template="{상품} 만드는 회사 이름 추천해줘. 기억에 남는 한글 이름으로",
)

prompt.format(상품="신발")

from langchain.chains import LLMChain
chain = LLMChain(llm=chat, prompt=prompt)

# chain.run("AI 여행 추천 서비스")
chain.run(상품="콜라")

#@title 4. ChatPromptTemplate & chain

from langchain.chat_models import ChatOpenAI
from langchain.prompts.chat import (
    ChatPromptTemplate,
    SystemMessagePromptTemplate,
    HumanMessagePromptTemplate,
)

chat = ChatOpenAI(temperature=0)

template="You are a helpful assisstant that tranlates {input_language} to {output_language}."
system_message_prompt = SystemMessagePromptTemplate.from_template(template)
human_template="{text}"
human_message_prompt = HumanMessagePromptTemplate.from_template(human_template)

chat_prompt = ChatPromptTemplate.from_messages([system_message_prompt, human_message_prompt])

chatchain = LLMChain(llm=chat, prompt=chat_prompt)
chatchain.run(input_language="English", output_language="Korean", text="Thanks god TGIF!")

#@title 5. Agents and Tools

from langchain.agents import load_tools
from langchain.agents import initialize_agent
from langchain.agents import AgentType

tools = load_tools(["serpapi", "llm-math"], llm=chat)
# tools = load_tools(["wikipedia", "llm-math"], llm=chat)

agent = initialize_agent(tools, llm=chat, agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION, verbose=True)

agent.run("한국의 'New Jeans' 그룹 멤버의 인원 수는? 오늘 2023년 11월 17일 현재 전체 멤버들의 나이의 합은?") # 한국의 'New Jeans' 그룹 멤버의 인원 수는? 오늘 2023년 5월 30일 현재 전체 멤버들의 나이의 합은 얼마인가요

len(agent.tools)

print(agent.tools[0].description)
print(agent.tools[1].description)

#@title 6. Memory
from langchain import ConversationChain

conversation = ConversationChain(llm=chat, verbose=True)
conversation.predict(input="인공지능에서 Transformer가 뭐야?")

conversation.predict(input="RNN하고 차이 설명해줘.")

conversation.predict(input="Seq2Seq 와도 차이점을 비교해줘.")

conversation.memory

#@title 7. Document Loaders
from langchain.document_loaders import WebBaseLoader

loader = WebBaseLoader(web_path="https://ko.wikipedia.org/wiki/NewJeans")

documents = loader.load()

from langchain.text_splitter import CharacterTextSplitter
text_splitter = CharacterTextSplitter(chunk_size=1000, chunk_overlap=0)
docs = text_splitter.split_documents(documents)
len(docs)

# 4096 token = 3000 English word

print(docs[-1].page_content)

#@title 8. Summarization
from langchain.chains.summarize import load_summarize_chain
chain = load_summarize_chain(chat, chain_type="map_reduce", verbose=True)
chain.run(docs[:-1])

#@title 9. Embeddings and VectorStore
from langchain.embeddings import HuggingFaceEmbeddings
from langchain.embeddings import OpenAIEmbeddings

# embeddings = OpenAIEmbeddings()
embeddings = HuggingFaceEmbeddings()

from langchain.indexes import VectorstoreIndexCreator
from langchain.vectorstores import FAISS

# from langchain.text_splitter import RecursiveCharacterTextSplitter
# text_splitter = RecursiveCharacterTextSplitter(chunk_size=1000, chunk_overlap=0)

index = VectorstoreIndexCreator(
    vectorstore_cls=FAISS,
    embedding=embeddings,
    # text_splitter=text_splitter,
    ).from_loaders([loader])

 

# 파일로 저장
index.vectorstore.save_local("faiss-nj")

index.query("뉴진스의 데뷔곡은?", llm=chat, verbose=True)

index.query("뉴진스의 데뷔 멤버는?", llm=chat, verbose=True)

index.query("멤버의 나이는?", llm=chat, verbose=True)

index.query("멤버의 나이는? (오늘은 2023년 11월 18일)", llm=chat, verbose=True)

#@title FAISS 벡터DB 디스크에서 불러오기
from langchain.indexes.vectorstore import VectorStoreIndexWrapper

fdb = FAISS.load_local("faiss-nj", embeddings)
index2 = VectorStoreIndexWrapper(vectorstore=fdb)

index2.query("뉴진스의 데뷔 멤버는?", llm=chat, verbose=True)

 

 

 

Memory

- ConversationBufferMemory : 대화 기록(기본)

- ConversationBufferWindowMemory : 마지막 n개의 대화만 기억

- Entity Memory : 개체에 대한 정보를 저장

- Conversation Knowledge Graph Memory: 개체의 triple 저장: (sam, 좋아하는 색, 파랑)

- ConversationSummaryMemory : 대화의 요약본을 저장

- ConversationSummaryBufferMemory : 대화 요약본 + 마지막 n토큰 기억

- ConversationTokenBufferMemory : 마지막 n토큰 기억

- VectorStore-Backed Memory : 벡터DB에 정보 저장

 

LIamaIndex vs Langchain

https://www.langchain.com/

LangChain

LangChain’s flexible abstractions and extensive toolkit unlocks developers to build context-aware, reasoning LLM applications.

www.langchain.com

!pip install --upgrade openai
!pip show openai | grep Version
!pip install llama-index pypdf
!pip install langchain

import openai
# Set up the OpenAI API client
openai.api_key = "sk-"

import os
openai_api_key = os.getenv('OPENAI_API_KEY', 'sk-')

import logging
import sys
logging.basicConfig(stream=sys.stdout, level=logging.INFO)
logging.getLogger().addHandler(logging.StreamHandler(stream=sys.stdout))

from langchain.agents import Tool
from langchain.chains.conversation.memory import ConversationBufferMemory
from langchain.chat_models import ChatOpenAI
from langchain.agents import initialize_agent

from llama_index import VectorStoreIndex, SimpleDirectoryReader

!mkdir data

documents = SimpleDirectoryReader("./data").load_data()
index = VectorStoreIndex.from_documents(documents=documents)

tools = [
    Tool(
        name="LlamaIndex",
        func=lambda q: str(index.as_query_engine().query(q)),
        description="useful for when you want to answer questions about the author. The input to this tool should be a complete english sentence.",
        return_direct=True,
    ),
]

# set Logging to DEBUG for more detailed outputs
memory = ConversationBufferMemory(memory_key="chat_history")
llm = ChatOpenAI(temperature=0, openai_api_key='sk-')
agent_executor = initialize_agent(
    tools, llm, agent="conversational-react-description", memory=memory
)

agent_executor.run(input="hi, i am bob")

agent_executor.run(input="What did the author do growing up?")

# try using SummaryIndex!
from langchain.llms import OpenAI
from langchain.llms import OpenAIChat
from langchain.agents import initialize_agent

from llama_index import SummaryIndex
from llama_index.langchain_helpers.memory_wrapper import GPTIndexChatMemory

index = SummaryIndex([])

# set Logging to DEBUG for more detailed outputs
# NOTE: you can also use a conversational chain

memory = GPTIndexChatMemory(
    index=index,
    memory_key="chat_history",
    query_kwargs={"response_mode": "compact"},
    # return_source returns source nodes instead of querying index
    return_source=True,
    # return_messages returns context in message format
    return_messages=True,
)
# llm = OpenAIChat(temperature=0)
llm=OpenAI(temperature=0, openai_api_key='sk-')
agent_executor = initialize_agent(
    [], llm, agent="conversational-react-description", memory=memory
)

agent_executor.run(input="hi, i am bob")

# NOTE: the query now calls the SummaryIndex memory module.
agent_executor.run(input="what's my name?")

 

 

!pip install --upgrade openai
!pip install llama-index pypdf
!pip install python-dotenv

# Environment Variables
import os
from dotenv import load_dotenv

openai_api_key = os.getenv('OPENAI_API_KEY', 'sk-')

import openai

# Set up the OpenAI API client
openai.api_key = "sk-"

# OpenAI_key = os.environ.get("OPENAI_API_KEY")
# print(OpenAI_key)

# OPENAI_API_KEY="sk-"
# openai.api_key=OPENAI_API_KEY

from langchain import OpenAI

from llama_index import SimpleDirectoryReader, ServiceContext, VectorStoreIndex
from llama_index import set_global_service_context
from llama_index.response.pprint_utils import pprint_response
from llama_index.tools import QueryEngineTool, ToolMetadata
from llama_index.query_engine import SubQuestionQueryEngine

# llm = OpenAI(temperature=0, model_name="text-davinci-003", max_tokens=-1)
llm = OpenAI(temperature=0, model_name="text-davinci-003", max_tokens=-1, openai_api_key="sk-")

service_context = ServiceContext.from_defaults(llm=llm)
set_global_service_context(service_context=service_context)

lyft_docs = SimpleDirectoryReader(input_files=["S-ECOPRO-20230814.pdf"]).load_data()
uber_docs = SimpleDirectoryReader(input_files=["S-INTERROJO-20230814.pdf"]).load_data()

print(f'Loaded lyft 10-K with {len(lyft_docs)} pages')
print(f'Loaded Uber 10-K with {len(uber_docs)} pages')

lyft_index = VectorStoreIndex.from_documents(lyft_docs)
uber_index = VectorStoreIndex.from_documents(uber_docs)

lyft_engine = lyft_index.as_query_engine(similarity_top_k=3)

uber_engine = uber_index.as_query_engine(similarity_top_k=3)

response = await lyft_engine.aquery('What is the summay of company in 2023?')

print(response)

response = await uber_engine.aquery('What is the summary of company in 2021?')

print(response)

 

 

 

 

 

 

 

 

 

 

 

 

 

 

Assistants -> function