Skip to main content
Open In Colab在 GitHub 上打开

Facebook 信使

此笔记本演示如何以可微调的格式从 Facebook 加载数据。总体步骤如下：

  1. 将您的 Messenger 数据下载到磁盘。
  2. 创建 Chat Loader 并调用loader.load()（或loader.lazy_load()） 执行转换。
  3. 可选使用merge_chat_runs按顺序合并来自同一发件人的消息，和/或map_ai_messages将指定发件人的邮件转换为 AIMessage 类。完成此作后，调用convert_messages_for_finetuning准备数据以进行微调。

完成此作后，您可以微调模型。为此，您需要完成以下步骤：

  1. 将您的消息上传到 OpenAI 并运行微调作业。
  2. 在您的 LangChain 应用程序中使用生成的模型！

让我们开始吧。

1. 下载数据

要下载您自己的 Messenger 数据，请按照此处的说明作。重要 - 确保以 JSON 格式（而不是 HTML）下载它们。

我们将在此 google drive 链接中托管一个示例转储，我们将在本演练中使用该链接。

# This uses some example data
import zipfile

import requests


def download_and_unzip(url: str, output_path: str = "file.zip") -> None:
    file_id = url.split("/")[-2]
    download_url = f"https://drive.google.com/uc?export=download&id={file_id}"

    response = requests.get(download_url)
    if response.status_code != 200:
        print("Failed to download the file.")
        return

    with open(output_path, "wb") as file:
        file.write(response.content)
        print(f"File {output_path} downloaded.")

    with zipfile.ZipFile(output_path, "r") as zip_ref:
        zip_ref.extractall()
        print(f"File {output_path} has been unzipped.")


# URL of the file to download
url = (
    "https://drive.google.com/file/d/1rh1s1o2i7B-Sk1v9o8KNgivLVGwJ-osV/view?usp=sharing"
)

# Download and unzip
download_and_unzip(url)
File file.zip downloaded.
File file.zip has been unzipped.

2. 创建 Chat Loader

我们有 2 种不同的FacebookMessengerChatLoader类，一个用于整个聊天目录，另一个用于加载单个文件。我们

directory_path = "./hogwarts"
from langchain_community.chat_loaders.facebook_messenger import (
    FolderFacebookMessengerChatLoader,
    SingleFileFacebookMessengerChatLoader,
)
loader = SingleFileFacebookMessengerChatLoader(
    path="./hogwarts/inbox/HermioneGranger/messages_Hermione_Granger.json",
)
chat_session = loader.load()[0]
chat_session["messages"][:3]
[HumanMessage(content="Hi Hermione! How's your summer going so far?", additional_kwargs={'sender': 'Harry Potter'}),
 HumanMessage(content="Harry! Lovely to hear from you. My summer is going well, though I do miss everyone. I'm spending most of my time going through my books and researching fascinating new topics. How about you?", additional_kwargs={'sender': 'Hermione Granger'}),
 HumanMessage(content="I miss you all too. The Dursleys are being their usual unpleasant selves but I'm getting by. At least I can practice some spells in my room without them knowing. Let me know if you find anything good in your researching!", additional_kwargs={'sender': 'Harry Potter'})]
loader = FolderFacebookMessengerChatLoader(
    path="./hogwarts",
)
chat_sessions = loader.load()
len(chat_sessions)
9

3. 准备微调

叫load()返回我们可以提取为人工消息的所有聊天消息。与聊天机器人交谈时，相对于真实对话，对话通常遵循更严格的交替对话模式。

您可以选择合并消息 “runs” （来自同一发件人的连续消息） 并选择发件人来表示 “AI”。经过微调的 LLM 将学习生成这些 AI 消息。

from langchain_community.chat_loaders.utils import (
    map_ai_messages,
    merge_chat_runs,
)
merged_sessions = merge_chat_runs(chat_sessions)
alternating_sessions = list(map_ai_messages(merged_sessions, "Harry Potter"))
# Now all of Harry Potter's messages will take the AI message class
# which maps to the 'assistant' role in OpenAI's training format
alternating_sessions[0]["messages"][:3]
[AIMessage(content="Professor Snape, I was hoping I could speak with you for a moment about something that's been concerning me lately.", additional_kwargs={'sender': 'Harry Potter'}),
 HumanMessage(content="What is it, Potter? I'm quite busy at the moment.", additional_kwargs={'sender': 'Severus Snape'}),
 AIMessage(content="I apologize for the interruption, sir. I'll be brief. I've noticed some strange activity around the school grounds at night. I saw a cloaked figure lurking near the Forbidden Forest last night. I'm worried someone may be plotting something sinister.", additional_kwargs={'sender': 'Harry Potter'})]

现在我们可以转换为 OpenAI 格式词典

from langchain_community.adapters.openai import convert_messages_for_finetuning
training_data = convert_messages_for_finetuning(alternating_sessions)
print(f"Prepared {len(training_data)} dialogues for training")
Prepared 9 dialogues for training
training_data[0][:3]
[{'role': 'assistant',
  'content': "Professor Snape, I was hoping I could speak with you for a moment about something that's been concerning me lately."},
 {'role': 'user',
  'content': "What is it, Potter? I'm quite busy at the moment."},
 {'role': 'assistant',
  'content': "I apologize for the interruption, sir. I'll be brief. I've noticed some strange activity around the school grounds at night. I saw a cloaked figure lurking near the Forbidden Forest last night. I'm worried someone may be plotting something sinister."}]

OpenAI 目前需要至少 10 个训练示例才能进行微调工作，尽管他们建议大多数任务在 50-100 个之间。由于我们只有 9 个聊天会话，因此我们可以对它们进行细分（可以选择进行一些重叠），以便每个训练示例都包含整个对话的一部分。

Facebook 聊天会话（每人 1 次）通常跨越多天和对话， 因此，无论如何，建模 Long-range dependency 可能并不那么重要。

# Our chat is alternating, we will make each datapoint a group of 8 messages,
# with 2 messages overlapping
chunk_size = 8
overlap = 2

training_examples = [
    conversation_messages[i : i + chunk_size]
    for conversation_messages in training_data
    for i in range(0, len(conversation_messages) - chunk_size + 1, chunk_size - overlap)
]

len(training_examples)
100

4. 微调模型

是时候微调模型了。确保您已openai安装 并已将OPENAI_API_KEY适当地

%pip install --upgrade --quiet  langchain-openai
import json
import time
from io import BytesIO

import openai

# We will write the jsonl file in memory
my_file = BytesIO()
for m in training_examples:
    my_file.write((json.dumps({"messages": m}) + "\n").encode("utf-8"))

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

# OpenAI audits each training file for compliance reasons.
# This make take a few minutes
status = openai.files.retrieve(training_file.id).status
start_time = time.time()
while status != "processed":
    print(f"Status=[{status}]... {time.time() - start_time:.2f}s", end="\r", flush=True)
    time.sleep(5)
    status = openai.files.retrieve(training_file.id).status
print(f"File {training_file.id} ready after {time.time() - start_time:.2f} seconds.")
File file-ULumAXLEFw3vB6bb9uy6DNVC ready after 0.00 seconds.

文件准备好后，就可以开始训练作业了。

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

在准备模型时喝杯茶。这可能需要一些时间！

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)
    job = openai.fine_tuning.jobs.retrieve(job.id)
    status = job.status
Status=[running]... 874.29s. 56.93s
print(job.fine_tuned_model)
ft:gpt-3.5-turbo-0613:personal::8QnAzWMr

5. 在 LangChain 中使用

您可以直接将生成的模型 ID 用于ChatOpenAImodel 类。

from langchain_openai import ChatOpenAI

model = ChatOpenAI(
    model=job.fine_tuned_model,
    temperature=1,
)
API 参考：ChatOpenAI
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate

prompt = ChatPromptTemplate.from_messages(
    [
        ("human", "{input}"),
    ]
)

chain = prompt | model | StrOutputParser()
for tok in chain.stream({"input": "What classes are you taking?"}):
    print(tok, end="", flush=True)
I'm taking Charms, Defense Against the Dark Arts, Herbology, Potions, Transfiguration, and Ancient Runes. How about you?