在这个笔记本中,我们将介绍如何使用函数调用对gpt-3.5-turbo进行微调。这里的主要用例是结构化数据提取。我们的主要重点是提炼GPT-4的输出,以帮助改进gpt-3.5-turbo的函数调用能力。
我们将从简单到高级逐个示例进行讲解:
%pip install llama-index-finetuning
%pip install llama-index-llms-openai
%pip install llama-index-finetuning-callbacks
%pip install llama-index-readers-file pymupdf
%pip install llama-index-program-openai
import nest_asyncio
nest_asyncio.apply()
import os
import openai
os.environ["OPENAI_API_KEY"] = "sk-..."
openai.api_key = os.environ["OPENAI_API_KEY"]
在本节中,我们将展示如何通过我们的低级Pydantic程序模块记录输入/输出。我们将使用该数据集对LLM进行微调。
在这里,我们定义了由 GPT-4 提供支持的函数调用程序,该程序将生成结构化输出到一个 Pydantic 对象(一个专辑)。
from llama_index.program.openai import OpenAIPydanticProgram
from pydantic import BaseModel
from llama_index.llms.openai import OpenAI
from llama_index.finetuning.callbacks import OpenAIFineTuningHandler
from llama_index.core.callbacks import CallbackManager
from typing import List
class Song(BaseModel):
"""歌曲的数据模型。"""
title: str
length_seconds: int
class Album(BaseModel):
"""专辑的数据模型。"""
name: str
artist: str
songs: List[Song]
finetuning_handler = OpenAIFineTuningHandler()
callback_manager = CallbackManager([finetuning_handler])
llm = OpenAI(model="gpt-4", callback_manager=callback_manager)
prompt_template_str = """\
生成一个示例专辑,包括艺术家和歌曲列表。\
以电影 {movie_name} 作为灵感。\
"""
program = OpenAIPydanticProgram.from_defaults(
output_cls=Album,
prompt_template_str=prompt_template_str,
llm=llm,
verbose=False,
)
我们定义一些示例电影名称作为输入,并通过函数调用程序记录输出。
# 注意:我们需要至少10部电影才能使用OpenAI微调
movie_names = [
"闪灵",
"无间道",
"泰坦尼克号",
"好家伙",
"风月情人",
"小鬼当家",
"铁笼狂怒",
"剪刀手爱德华",
"全面回忆",
"幽灵",
"震颤",
"机械战警",
"洛基5",
]
from tqdm.notebook import tqdm
for movie_name in tqdm(movie_names):
output = program(movie_name=movie_name)
print(output.json())
0%| | 0/13 [00:00<?, ?it/s]
{"name": "The Shining", "artist": "Various Artists", "songs": [{"title": "Main Title", "length_seconds": 180}, {"title": "Opening Credits", "length_seconds": 120}, {"title": "The Overlook Hotel", "length_seconds": 240}, {"title": "Redrum", "length_seconds": 150}, {"title": "Here's Johnny!", "length_seconds": 200}]} {"name": "The Departed Soundtrack", "artist": "Various Artists", "songs": [{"title": "Gimme Shelter", "length_seconds": 272}, {"title": "Comfortably Numb", "length_seconds": 383}, {"title": "I'm Shipping Up to Boston", "length_seconds": 166}, {"title": "Sweet Dreams (Are Made of This)", "length_seconds": 216}, {"title": "I'm Shipping Up to Boston (Instrumental)", "length_seconds": 166}, {"title": "The Departed Tango", "length_seconds": 123}, {"title": "Thief's Theme", "length_seconds": 201}, {"title": "Well Well Well", "length_seconds": 126}, {"title": "Comfortably Numb (Live)", "length_seconds": 383}, {"title": "Sail On, Sailor", "length_seconds": 181}]} {"name": "Titanic Soundtrack", "artist": "James Horner", "songs": [{"title": "My Heart Will Go On", "length_seconds": 273}, {"title": "Rose", "length_seconds": 120}, {"title": "Hymn to the Sea", "length_seconds": 365}, {"title": "Southampton", "length_seconds": 180}, {"title": "Take Her to Sea, Mr. Murdoch", "length_seconds": 150}]} {"name": "Goodfellas Soundtrack", "artist": "Various Artists", "songs": [{"title": "Rags to Riches", "length_seconds": 180}, {"title": "Gimme Shelter", "length_seconds": 270}, {"title": "Layla", "length_seconds": 270}, {"title": "Jump into the Fire", "length_seconds": 240}, {"title": "Atlantis", "length_seconds": 180}, {"title": "Beyond the Sea", "length_seconds": 180}, {"title": "Sunshine of Your Love", "length_seconds": 240}, {"title": "Mannish Boy", "length_seconds": 240}, {"title": "Layla (Piano Exit)", "length_seconds": 120}]} {"name": "Pretty Woman Soundtrack", "artist": "Various Artists", "songs": [{"title": "Oh, Pretty Woman", "length_seconds": 178}, {"title": "King of Wishful Thinking", "length_seconds": 253}, {"title": "It Must Have Been Love", "length_seconds": 250}, {"title": "Show Me Your Soul", "length_seconds": 285}, {"title": "No Explanation", "length_seconds": 244}]} {"name": "Home Alone Soundtrack", "artist": "John Williams", "songs": [{"title": "Somewhere in My Memory", "length_seconds": 180}, {"title": "Holiday Flight", "length_seconds": 120}, {"title": "The House", "length_seconds": 150}, {"title": "Star of Bethlehem", "length_seconds": 135}, {"title": "Setting the Trap", "length_seconds": 165}, {"title": "The Attack on the House", "length_seconds": 200}, {"title": "Mom Returns and Finale", "length_seconds": 240}]} {"name": "Caged Fury", "artist": "The Fury Band", "songs": [{"title": "Caged Fury", "length_seconds": 240}, {"title": "Prison Break", "length_seconds": 180}, {"title": "Behind Bars", "length_seconds": 210}, {"title": "Escape Plan", "length_seconds": 195}, {"title": "Fight for Freedom", "length_seconds": 220}]} {"name": "Edward Scissorhands Soundtrack", "artist": "Danny Elfman", "songs": [{"title": "Introduction", "length_seconds": 120}, {"title": "Ice Dance", "length_seconds": 180}, {"title": "Edwardo the Barber", "length_seconds": 150}, {"title": "The Grand Finale", "length_seconds": 240}]} {"name": "Total Recall", "artist": "Various Artists", "songs": [{"title": "Recall", "length_seconds": 240}, {"title": "Mars", "length_seconds": 180}, {"title": "Memory", "length_seconds": 210}, {"title": "Rebellion", "length_seconds": 300}, {"title": "Escape", "length_seconds": 270}]} {"name": "Ghost", "artist": "Various Artists", "songs": [{"title": "Unchained Melody", "length_seconds": 218}, {"title": "Oh My Love", "length_seconds": 156}, {"title": "Ditto's Theme", "length_seconds": 92}, {"title": "Love Inside", "length_seconds": 180}, {"title": "Ghostly Encounter", "length_seconds": 120}]} {"name": "Tremors Soundtrack", "artist": "Various Artists", "songs": [{"title": "Main Theme", "length_seconds": 180}, {"title": "Graboids Attack", "length_seconds": 240}, {"title": "Val and Earl's Theme", "length_seconds": 200}, {"title": "Burt's Arsenal", "length_seconds": 220}, {"title": "Nest of the Graboids", "length_seconds": 190}]} {"name": "RoboCop: The Soundtrack", "artist": "Various Artists", "songs": [{"title": "Main Theme", "length_seconds": 180}, {"title": "Murphy's Death", "length_seconds": 240}, {"title": "RoboCop's Training", "length_seconds": 210}, {"title": "ED-209", "length_seconds": 195}, {"title": "Clarence Boddicker", "length_seconds": 220}, {"title": "RoboCop Saves the Day", "length_seconds": 240}, {"title": "RoboCop's Theme", "length_seconds": 180}]} {"name": "Rocky V", "artist": "Various Artists", "songs": [{"title": "Measure of a Man", "length_seconds": 240}, {"title": "Can't Stop the Fire", "length_seconds": 210}, {"title": "Go for It!", "length_seconds": 180}, {"title": "Take You Back (Home Sweet Home)", "length_seconds": 200}, {"title": "The Measure of a Man (Reprise)", "length_seconds": 120}]}
finetuning_handler.save_finetuning_events("mock_finetune_songs.jsonl")
Wrote 14 examples to mock_finetune_songs.jsonl
!cat mock_finetune_songs.jsonl
现在我们定义一个微调引擎,并在模拟数据集上进行微调。
from llama_index.finetuning import OpenAIFinetuneEngine
finetune_engine = OpenAIFinetuneEngine(
"gpt-3.5-turbo",
"mock_finetune_songs.jsonl",
# start_job_id="<start-job-id>" # 如果你有一个现有的作业,可以在这里指定id
validate_json=False, # openai验证json代码尚不支持函数调用
)
finetune_engine.finetune()
finetune_engine.get_current_job()
<FineTuningJob fine_tuning.job id=ftjob-uJ9kQ9pI0p0YNatBDxF3VITv at 0x172a5c9a0> JSON: { "object": "fine_tuning.job", "id": "ftjob-uJ9kQ9pI0p0YNatBDxF3VITv", "model": "gpt-3.5-turbo-0613", "created_at": 1696463378, "finished_at": 1696463749, "fine_tuned_model": "ft:gpt-3.5-turbo-0613:llamaindex::8660TXqx", "organization_id": "org-1ZDAvajC6v2ZtAP9hLEIsXRz", "result_files": [ "file-Hbpw15BAwyf3e4HK5Z9g4IK2" ], "status": "succeeded", "validation_file": null, "training_file": "file-MNh7snhv0triDIhsrErokSMY", "hyperparameters": { "n_epochs": 7 }, "trained_tokens": 22834, "error": null }
我们获得了经过微调的LLM,并将其与Pydantic程序一起使用。
ft_llm = finetune_engine.get_finetuned_model(temperature=0.3)
ft_program = OpenAIPydanticProgram.from_defaults(
output_cls=Album,
prompt_template_str=prompt_template_str,
llm=ft_llm,
verbose=False,
)
ft_program(movie_name="Goodfellas")
Album(name='Goodfellas Soundtrack', artist='Various Artists', songs=[Song(title='Rags to Riches', length_seconds=180), Song(title='Gimme Shelter', length_seconds=270), Song(title='Layla', length_seconds=270), Song(title='Jump into the Fire', length_seconds=240), Song(title='Atlantis', length_seconds=180), Song(title='Beyond the Sea', length_seconds=180), Song(title='Sunshine of Your Love', length_seconds=240), Song(title='Mannish Boy', length_seconds=240), Song(title='Layla (Piano Exit)', length_seconds=120)])
函数调用的一个用例是通过RAG系统获取结构化输出。
在这里,我们展示如何创建一个训练数据集,其中包括上下文增强输入和未结构化文档上的结构化输出。然后,我们可以对LLM进行微调,并将其插入到RAG系统中,以执行检索和输出提取。
!mkdir data && wget --user-agent "Mozilla" "https://arxiv.org/pdf/2307.09288.pdf" -O "data/llama2.pdf"
--2023-10-04 23:46:36-- https://arxiv.org/pdf/2307.09288.pdf Resolving arxiv.org (arxiv.org)... 128.84.21.199 Connecting to arxiv.org (arxiv.org)|128.84.21.199|:443... connected. HTTP request sent, awaiting response... 200 OK Length: 13661300 (13M) [application/pdf] Saving to: ‘data/llama2.pdf’ data/llama2.pdf 100%[===================>] 13.03M 229KB/s in 45s 2023-10-04 23:47:25 (298 KB/s) - ‘data/llama2.pdf’ saved [13661300/13661300]
from pydantic import Field
from typing import List
class Citation(BaseModel):
"""引文类。"""
author: str = Field(
..., description="推断出的第一作者(通常是姓氏)"
)
year: int = Field(..., description="推断出的年份")
desc: str = Field(
...,
description=(
"从作者被引用的作品的文本中推断出的描述"
),
)
class Response(BaseModel):
"""作者引文列表。
从非结构化文本中提取。
"""
citations: List[Citation] = Field(
...,
description=(
"作者引文列表(按作者、年份和描述组织)。"
),
)
import pandas as pd
# 读取数据
data = pd.read_csv('data.csv')
# 显示数据的前几行
data.head()
from llama_index.readers.file import PyMuPDFReader
from llama_index.core import Document
from llama_index.core.node_parser import SentenceSplitter
from pathlib import Path
loader = PyMuPDFReader()
docs0 = loader.load(file_path=Path("./data/llama2.pdf"))
doc_text = "\n\n".join([d.get_content() for d in docs0])
metadata = {
"paper_title": "Llama 2: Open Foundation and Fine-Tuned Chat Models"
}
docs = [Document(text=doc_text, metadata=metadata)]
chunk_size = 1024
node_parser = SentenceSplitter(chunk_size=chunk_size)
nodes = node_parser.get_nodes_from_documents(docs)
len(nodes)
89
from llama_index.core import Settings
finetuning_handler = OpenAIFineTuningHandler()
callback_manager = CallbackManager([finetuning_handler])
Settings.chunk_size = chunk_size
gpt_4_llm = OpenAI(
model="gpt-4-0613", temperature=0.3, callback_manager=callback_manager
)
gpt_35_llm = OpenAI(
model="gpt-3.5-turbo-0613",
temperature=0.3,
callback_manager=callback_manager,
)
eval_llm = OpenAI(model="gpt-4-0613", temperature=0)
在这里,我们展示如何在这些非结构化的块/节点上生成一个训练数据集。
我们生成问题来提取不同上下文中的引用。我们通过一个GPT-4 RAG管道运行这些问题,提取结构化输出,并记录输入/输出。
# 设置数据集生成器
from llama_index.core.evaluation import DatasetGenerator
from llama_index.core import SummaryIndex
from llama_index.core import PromptTemplate
from tqdm.notebook import tqdm
from tqdm.asyncio import tqdm_asyncio
fp = open("data/qa_pairs.jsonl", "w")
question_gen_prompt = PromptTemplate(
"""
{query_str}
Context:
{context_str}
Questions:
"""
)
question_gen_query = """\
给定一篇研究论文的片段。它包含引用。
请从文本中生成关于这些引用的问题。
例如,以下是一些示例问题:
哪些引用对应于变压器模型的相关工作?
告诉我关于推进RLHF的作者。
你能告诉我所有计算机视觉作品对应的引用吗?\
"""
qr_pairs = []
node_questions_tasks = []
for idx, node in enumerate(nodes[:39]):
num_questions = 1 # 更改此数字以增加节点数量
dataset_generator = DatasetGenerator(
[node],
question_gen_query=question_gen_query,
text_question_template=question_gen_prompt,
llm=eval_llm,
metadata_mode="all",
num_questions_per_chunk=num_questions,
)
task = dataset_generator.agenerate_questions_from_nodes(num=num_questions)
node_questions_tasks.append(task)
node_questions_lists = await tqdm_asyncio.gather(*node_questions_tasks)
node_questions_lists
from llama_index.core import VectorStoreIndex
gpt4_index = VectorStoreIndex(nodes=nodes)
gpt4_query_engine = gpt4_index.as_query_engine(
output_cls=Response, similarity_top_k=1, llm=gpt_4_llm
)
from json import JSONDecodeError
for idx, node in enumerate(tqdm(nodes[:39])):
node_questions_0 = node_questions_lists[idx]
for question in node_questions_0:
try:
# 注意:我们不需要使用response,事件通过fine-tuning处理程序记录
gpt4_query_engine.query(question)
except Exception as e:
print(f"问题 {question} 出错, {repr(e)}")
pass
0%| | 0/39 [00:00<?, ?it/s]
Error for question Which citations are referred to in the discussion about safety investigations into pretraining data and pretrained models?, ValidationError(model='Response', errors=[{'loc': ('__root__',), 'msg': 'Expecting value: line 1 column 1 (char 0)', 'type': 'value_error.jsondecode', 'ctx': {'msg': 'Expecting value', 'doc': 'Empty Response', 'pos': 0, 'lineno': 1, 'colno': 1}}])
finetuning_handler.save_finetuning_events("llama2_citation_events.jsonl")
Wrote 83 examples to llama2_citation_events.jsonl
我们开始对生成的数据集进行微调。
from llama_index.finetuning import OpenAIFinetuneEngine
finetune_engine = OpenAIFinetuneEngine(
"gpt-3.5-turbo",
"llama2_citation_events.jsonl",
# start_job_id="<start-job-id>" # 如果你有一个现有的作业,可以在这里指定id
validate_json=False, # openai验证json代码尚不支持函数调用
)
finetune_engine.finetune()
finetune_engine.get_current_job()
<FineTuningJob fine_tuning.job id=ftjob-ATYm4yZHP1QvXs1wx85Ix79F at 0x1752b6b60> JSON: { "object": "fine_tuning.job", "id": "ftjob-ATYm4yZHP1QvXs1wx85Ix79F", "model": "gpt-3.5-turbo-0613", "created_at": 1696497663, "finished_at": 1696498092, "fine_tuned_model": "ft:gpt-3.5-turbo-0613:llamaindex::86EwPw83", "organization_id": "org-1ZDAvajC6v2ZtAP9hLEIsXRz", "result_files": [ "file-wabcIIxjLqvhqOVohf4qSmE7" ], "status": "succeeded", "validation_file": null, "training_file": "file-WbYcsinIbH8vyCAstcoFEr92", "hyperparameters": { "n_epochs": 3 }, "trained_tokens": 132678, "error": null }
让我们将经过微调的LLM插入到一个完整的RAG pipeline中,以输出结构化的结果。
ft_llm = finetune_engine.get_finetuned_model(temperature=0.3)
from llama_index.core import VectorStoreIndex
vector_index = VectorStoreIndex(nodes=nodes)
query_engine = vector_index.as_query_engine(
output_cls=Response, similarity_top_k=1, llm=ft_llm
)
# 将基线设置为
base_index = VectorStoreIndex(nodes=nodes)
base_query_engine = base_index.as_query_engine(
output_cls=Response, similarity_top_k=1, llm=gpt_35_llm
)
query_str = """\
用于衡量Llama 2真实性的引用是哪个?\
"""
response = query_engine.query(query_str)
print(str(response))
{"citations": [{"author": "Lin et al.", "year": 2021, "desc": "TruthfulQA, used for LLM hallucinations to measure whether a language model is truthful in generating answers to questions while being informative at the same time."}]}
base_response = base_query_engine.query(query_str)
print(str(base_response))
{"citations": [{"author": "Lin et al.", "year": 2021, "desc": "TruthfulQA"}]}
# 查看源代码
print(response.source_nodes[0].get_content())
# 作为参考,请查看GPT-4的响应
gpt4_response = gpt4_query_engine.query(query_str)
print(str(gpt4_response))
{"citations": [{"author": "Lin et al.", "year": 2021, "desc": "TruthfulQA, used for LLM hallucinations to measure whether a language model is truthful in generating answers to questions while being informative at the same time."}]}