LangSmith LLM 运行

本笔记本演示了如何直接从 LangSmith 的 LLM 运行中加载数据并用这些数据微调模型。该过程简单，包含 3 个步骤。

1. 选择用于训练的 LLM 运行。
2. 使用 LangSmithRunChatLoader 将运行加载为聊天会话。
3. 微调你的模型。

然后你就可以在你的 LangChain 应用中使用微调后的模型了。

在深入了解之前，我们先安装所需的先决条件。

先决条件

请确保你已安装 langchain >= 0.0.311，并已使用你的 LangSmith API 密钥配置了环境。

%pip install --upgrade --quiet  langchain langchain-openai

import os
import uuid

uid = uuid.uuid4().hex[:6]
project_name = f"Run Fine-tuning Walkthrough {uid}"
os.environ["LANGSMITH_TRACING"] = "true"
os.environ["LANGSMITH_API_KEY"] = "YOUR API KEY"
os.environ["LANGSMITH_PROJECT"] = project_name

1. 选择运行

第一步是选择要进行微调的运行。一个常见的场景是选择在收到积极用户反馈的跟踪中的 LLM 运行。你可以在LangSmith Cookbook 和文档中找到相关示例。

为了本教程的目的，我们将为你生成一些运行供此处使用。让我们尝试微调一个简单的函数调用链。

from enum import Enum

from pydantic import BaseModel, Field


class Operation(Enum):
    add = "+"
    subtract = "-"
    multiply = "*"
    divide = "/"


class Calculator(BaseModel):
    """A calculator function"""

    num1: float
    num2: float
    operation: Operation = Field(..., description="+,-,*,/")

    def calculate(self):
        if self.operation == Operation.add:
            return self.num1 + self.num2
        elif self.operation == Operation.subtract:
            return self.num1 - self.num2
        elif self.operation == Operation.multiply:
            return self.num1 * self.num2
        elif self.operation == Operation.divide:
            if self.num2 != 0:
                return self.num1 / self.num2
            else:
                return "Cannot divide by zero"

from pprint import pprint

from langchain_core.utils.function_calling import convert_pydantic_to_openai_function
from pydantic import BaseModel

openai_function_def = convert_pydantic_to_openai_function(Calculator)
pprint(openai_function_def)

API 参考：convert_pydantic_to_openai_function

{'description': 'A calculator function',
 'name': 'Calculator',
 'parameters': {'description': 'A calculator function',
                'properties': {'num1': {'title': 'Num1', 'type': 'number'},
                               'num2': {'title': 'Num2', 'type': 'number'},
                               'operation': {'allOf': [{'description': 'An '
                                                                       'enumeration.',
                                                        'enum': ['+',
                                                                 '-',
                                                                 '*',
                                                                 '/'],
                                                        'title': 'Operation'}],
                                             'description': '+,-,*,/'}},
                'required': ['num1', 'num2', 'operation'],
                'title': 'Calculator',
                'type': 'object'}}

from langchain_core.output_parsers.openai_functions import PydanticOutputFunctionsParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_openai import ChatOpenAI

prompt = ChatPromptTemplate.from_messages(
    [
        ("system", "You are an accounting assistant."),
        ("user", "{input}"),
    ]
)
chain = (
    prompt
    | ChatOpenAI().bind(functions=[openai_function_def])
    | PydanticOutputFunctionsParser(pydantic_schema=Calculator)
    | (lambda x: x.calculate())
)

API 参考：PydanticOutputFunctionsParser | ChatPromptTemplate | ChatOpenAI

math_questions = [
    "What's 45/9?",
    "What's 81/9?",
    "What's 72/8?",
    "What's 56/7?",
    "What's 36/6?",
    "What's 64/8?",
    "What's 12*6?",
    "What's 8*8?",
    "What's 10*10?",
    "What's 11*11?",
    "What's 13*13?",
    "What's 45+30?",
    "What's 72+28?",
    "What's 56+44?",
    "What's 63+37?",
    "What's 70-35?",
    "What's 60-30?",
    "What's 50-25?",
    "What's 40-20?",
    "What's 30-15?",
]
results = chain.batch([{"input": q} for q in math_questions], return_exceptions=True)

加载未出错的运行

现在我们可以选择成功的运行进行微调。

from langsmith.client import Client

client = Client()

successful_traces = {
    run.trace_id
    for run in client.list_runs(
        project_name=project_name,
        execution_order=1,
        error=False,
    )
}

llm_runs = [
    run
    for run in client.list_runs(
        project_name=project_name,
        run_type="llm",
    )
    if run.trace_id in successful_traces
]

2. 准备数据

现在我们可以创建一个 LangSmithRunChatLoader 实例，并使用其 lazy_load() 方法加载聊天会话。

from langchain_community.chat_loaders.langsmith import LangSmithRunChatLoader

loader = LangSmithRunChatLoader(runs=llm_runs)

chat_sessions = loader.lazy_load()

API 参考：LangSmithRunChatLoader

加载聊天会话后，将其转换为适合微调的格式。

from langchain_community.adapters.openai import convert_messages_for_finetuning

training_data = convert_messages_for_finetuning(chat_sessions)

API 参考：convert_messages_for_finetuning

3. 微调模型

现在，使用 OpenAI 库启动微调过程。

import json
import time
from io import BytesIO

import openai

my_file = BytesIO()
for dialog in training_data:
    my_file.write((json.dumps({"messages": dialog}) + "\n").encode("utf-8"))

my_file.seek(0)
training_file = openai.files.create(file=my_file, purpose="fine-tune")

job = openai.fine_tuning.jobs.create(
    training_file=training_file.id,
    model="gpt-3.5-turbo",
)

# Wait for the fine-tuning to complete (this may take some time)
status = openai.fine_tuning.jobs.retrieve(job.id).status
start_time = time.time()
while status != "succeeded":
    print(f"Status=[{status}]... {time.time() - start_time:.2f}s", end="\r", flush=True)
    time.sleep(5)
    status = openai.fine_tuning.jobs.retrieve(job.id).status

# Now your model is fine-tuned!

Status=[running]... 349.84s. 17.72s

4. 在 LangChain 中使用

微调完成后，在你的 LangChain 应用中将生成的模型 ID 与 ChatOpenAI 模型类一起使用。

# Get the fine-tuned model ID
job = openai.fine_tuning.jobs.retrieve(job.id)
model_id = job.fine_tuned_model

# Use the fine-tuned model in LangChain
from langchain_openai import ChatOpenAI

model = ChatOpenAI(
    model=model_id,
    temperature=1,
)

API 参考：ChatOpenAI

(prompt | model).invoke({"input": "What's 56/7?"})

AIMessage(content='Let me calculate that for you.')

现在，你已成功使用 LangSmith LLM 运行数据微调了一个模型！