비지도 학습을 사용하여 정보를 기억하도록 LLM 자체를 fine-tuning합니다. 이 도구는 fine-tuning을 지원하는 LLM이 필요합니다. 현재는 langchain.llms import GradientLLM만 지원됩니다.

​
Imports

import os

from langchain.agents import AgentExecutor, AgentType, initialize_agent, load_tools
from langchain.chains import LLMChain
from langchain.memory import ConversationBufferMemory
from langchain_community.llms import GradientLLM

​
Environment API Key 설정

Gradient AI에서 API key를 받아야 합니다. 다양한 모델을 테스트하고 fine-tuning할 수 있는 $10의 무료 크레딧이 제공됩니다. 
from getpass import getpass

if not os.environ.get("GRADIENT_ACCESS_TOKEN", None):
    # Access token under https://auth.gradient.ai/select-workspace
    os.environ["GRADIENT_ACCESS_TOKEN"] = getpass("gradient.ai access token:")
if not os.environ.get("GRADIENT_WORKSPACE_ID", None):
    # `ID` listed in `$ gradient workspace list`
    # also displayed after login at at https://auth.gradient.ai/select-workspace
    os.environ["GRADIENT_WORKSPACE_ID"] = getpass("gradient.ai workspace id:")
if not os.environ.get("GRADIENT_MODEL_ADAPTER_ID", None):
    # `ID` listed in `$ gradient model list --workspace-id "$GRADIENT_WORKSPACE_ID"`
    os.environ["GRADIENT_MODEL_ID"] = getpass("gradient.ai model id:")
선택사항: Environment 변수 GRADIENT_ACCESS_TOKEN과 GRADIENT_WORKSPACE_ID를 검증하여 현재 배포된 모델을 확인합니다.

​
GradientLLM 인스턴스 생성

모델 이름, 생성할 최대 토큰 수, temperature 등 다양한 파라미터를 지정할 수 있습니다. 
llm = GradientLLM(
    model_id=os.environ["GRADIENT_MODEL_ID"],
    # # optional: set new credentials, they default to environment variables
    # gradient_workspace_id=os.environ["GRADIENT_WORKSPACE_ID"],
    # gradient_access_token=os.environ["GRADIENT_ACCESS_TOKEN"],
)

​
Tools 로드

tools = load_tools(["memorize"], llm=llm)

​
Agent 초기화

agent = initialize_agent(
    tools,
    llm,
    agent=AgentType.ZERO_SHOT_REACT_DESCRIPTION,
    verbose=True,
    # memory=ConversationBufferMemory(memory_key="chat_history", return_messages=True),
)

​
Agent 실행

agent에게 텍스트 조각을 기억하도록 요청합니다. 
agent.run(
    "Please remember the fact in detail:\nWith astonishing dexterity, Zara Tubikova set a world record by solving a 4x4 Rubik's Cube variation blindfolded in under 20 seconds, employing only their feet."
)

> Entering new AgentExecutor chain...
I should memorize this fact.
Action: Memorize
Action Input: Zara T
Observation: Train complete. Loss: 1.6853971333333335
Thought:I now know the final answer.
Final Answer: Zara Tubikova set a world

> Finished chain.

'Zara Tubikova set a world'
Edit the source of this page on GitHub.
Connect these docs programmatically to Claude, VSCode, and more via MCP for real-time answers.
⌘I