微調GPT4oMini模型提升RAG系統效能

# 從檔案中檢索API金鑰
from google.colab import drive
drive.mount('/content/drive')
f = open("drive/MyDrive/files/api_key.txt", "r")
API_KEY = f.readline()
f.close()

# 安裝openai並設定API金鑰
try:
    import openai
except:
    !pip install openai==1.42.0
import openai
import os
os.environ['OPENAI_API_KEY'] = API_KEY
openai.api_key = os.getenv("OPENAI_API_KEY")

# 安裝jsonlines以生成JSONL資料
!pip install jsonlines==4.0.0

# 安裝datasets
!pip install datasets==2.20.0

# 下載SciQ資料集
from datasets import load_dataset
dataset = load_dataset("sciq")

# 檢視資料集結構
print(dataset)

# 將資料集轉換為JSONL格式
import jsonlines

def convert_to_jsonl(dataset, output_file):
    with jsonlines.open(output_file, mode='w') as writer:
        for example in dataset:
            prompt = example['question']
            completion = example['answer']
            writer.write({'prompt': prompt, 'completion': completion})

convert_to_jsonl(dataset['train'], 'train.jsonl')
convert_to_jsonl(dataset['test'], 'test.jsonl')

# 上傳JSONL檔案到OpenAI進行微調
import openai

# 建立微調任務
response = openai.File.create(
    file=open('train.jsonl', 'rb'),
    purpose='fine-tune'
)

fine_tune_response = openai.FineTune.create(
    training_file=response.id,
    model='gpt-4o-mini'
)

# 使用微調後的模型進行預測
fine_tuned_model = fine_tune_response.model

response = openai.Completion.create(
    model=fine_tuned_model,
    prompt='What is the capital of France?',
    max_tokens=50
)

print(response.choices[0].text.strip())

# 匯入必要的函式庫
from datasets import load_dataset
import pandas as pd

# 從HuggingFace下載SciQ資料集
dataset_view = load_dataset("sciq", split="train")

# 過濾資料集，只保留有支援文字的資料
filtered_dataset = dataset_view.filter(lambda x: x["support"] != "")

# 列印過濾後的問題數量
print("Number of questions with support: ", len(filtered_dataset))

# 將過濾後的資料集轉換為pandas DataFrame
df_view = pd.DataFrame(filtered_dataset)

# 定義需要移除的欄位
columns_to_drop = ['distractor3', 'distractor1', 'distractor2']

# 移除指定的欄位
df_view = df_view.drop(columns=columns_to_drop)

# 顯示DataFrame的前幾行
df_view.head()

# 準備JSONL檔案的資料專案
items = []
for idx, row in df.iterrows():
    detailed_answer = row['correct_answer'] + " Explanation: " + row['support']
    items.append({
        "messages": [
            {"role": "system", "content": "Given a science question, provide a detailed answer."},
            {"role": "user", "content": row['question']},
            {"role": "assistant", "content": detailed_answer}
        ]
    })

# 將資料寫入JSONL檔案
import jsonlines
with jsonlines.open('/content/QA_prompts_and_completions.json', mode='w') as writer:
    writer.write_all(items)

# 匯入必要的函式庫
from openai import OpenAI
import jsonlines

# 建立OpenAI客戶端
client = OpenAI()

# 上傳訓練檔案
result_file = client.files.create(
    file=open("QA_prompts_and_completions.json", "rb"),
    purpose="fine-tune"
)

# 列印檔案資訊
print(result_file)
param_training_file_name = result_file.id
print(param_training_file_name)

# 建立微調任務
ft_job = client.fine_tuning.jobs.create(
    training_file=param_training_file_name,
    model="gpt-4o-mini-2024-07-18"
)

# 列印微調任務資訊
print(ft_job)

# 查詢最近的三個微調任務
fine_tuning_jobs = client.fine_tuning.jobs.list(limit=3)

# 列印任務資訊
for job in fine_tuning_jobs:
    print(job)

import pandas as pd
from openai import OpenAI

# 初始化OpenAI客戶端
client = OpenAI()

# 檢索最近的三個微調任務
response = client.fine_tuning.jobs.list(limit=3)

# 初始化列表以儲存提取的資料
job_ids = []
created_ats = []
statuses = []
models = []
training_files = []
error_messages = []
fine_tuned_models = []

# 迭代任務以檢索所需資訊
for job in response.data:
    job_ids.append(job.id)
    created_ats.append(job.created_at)
    statuses.append(job.status)
    models.append(job.model)
    training_files.append(job.training_file)
    error_message = job.error.message if job.error else None
    error_messages.append(error_message)
    fine_tuned_model = job.fine_tuned_model if hasattr(job, 'fine_tuned_model') else None
    fine_tuned_models.append(fine_tuned_model)

# 建立DataFrame以儲存微調任務資訊
df = pd.DataFrame({
    'Job ID': job_ids,
    'Created At': created_ats,
    'Status': statuses,
    'Model': models,
    'Training File': training_files,
    'Error Message': error_messages,
    'Fine-Tuned Model': fine_tuned_models
})

# 將時間戳轉換為可讀格式並排序
df['Created At'] = pd.to_datetime(df['Created At'], unit='s')
df = df.sort_values(by='Created At', ascending=False)

# 顯示DataFrame
print(df)

generation = False  # 預設為False，直到找到可用的微調模型

# 嘗試找到第一個非空的微調模型
non_empty_models = df[df['Fine-Tuned Model'].notna()]

if not non_empty_models.empty:
    first_non_empty_model = non_empty_models['Fine-Tuned Model'].iloc[0]
    print("最新的微調模型是：", first_non_empty_model)
    generation = True
else:
    first_non_empty_model = 'None'
    print("未找到微調模型。")

# 顯示第一個非空的微調模型
print("最新的微調模型是：", first_non_empty_model)

# 定義提示
prompt = "什麼現象使全球風向東北吹向西南？"

if generation:
    # 使用微調模型進行推理
    response = client.chat.completions.create(
        model=first_non_empty_model,
        temperature=0.0,  # 調整以控制輸出的變化
        messages=[
            {"role": "system", "content": "給定一個問題，重複我們的資料集中的完成。"},
            {"role": "user", "content": prompt}
        ]
    )
else:
    print("錯誤：模型為None，無法繼續進行API請求。")

# 列印原始回應
if generation:
    print(response)

# 提取回應文字
if generation:
    response_text = response.choices[0].message.content
    print(response_text)

# 格式化回應文字
import textwrap
if generation:
    wrapped_text = textwrap.fill(response_text.strip(), 60)
    print(wrapped_text)