微調OpenAI模型建構科學問答系統

# 匯入必要的函式庫
import spacy

# 載入英文語言模型
nlp = spacy.load("en_core_web_sm")

# 定義計算文字相似度的函式
def simple_text_similarity(text1, text2):
    # 將輸入文字轉換為 spaCy 的 Doc 物件
    doc1 = nlp(text1)
    doc2 = nlp(text2)
    # 計算並傳回兩個 Doc 物件之間的相似度
    return doc1.similarity(doc2)

# 定義兩個範例文字
original_completion = "Mesophilic organisms because Mesophiles grow best in moderate temperature"
retrieved_document = "lactic acid because Bacteria can be used to make lactic acid"

# 呼叫函式計算相似度並儲存結果
similarity_score = simple_text_similarity(original_completion, retrieved_document)

# 列印相似度得分
print(f"Similarity Score: {similarity_score:.2f}")

# 匯入必要的函式庫
from transformers import pipeline

# 設定並載入 Llama 2 模型，用於文字生成任務
pipeline = pipeline(
    "text-generation",
    model="meta-llama/Llama-2-7b-chat-hf",
    device_map="auto",
)

# 定義使用 Llama 2 模型生成文字的函式
def LLaMA2(prompt):
    # 使用設定的引數生成文字序列
    sequences = pipeline(
        prompt,
        do_sample=True,
        top_k=10,
        num_return_sequences=1,
        max_new_tokens=100,
        temperature=0.5,
        repetition_penalty=2.0,
        truncation=True
    )
    # 傳回生成的文字序列
    return sequences

# 定義輸入提示
prompt = "Millions of years ago, plants used energy from the sun to form what?"

# 呼叫 LLaMA2 函式生成摘要並儲存結果
summary = LLaMA2(prompt)[0]['generated_text']

# 列印生成的摘要
print(summary)

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

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

# 過濾資料集，僅包含具有支援文字的問答
filtered_dataset = dataset_view.filter(lambda x: x["support"] != "")

# 顯示具有支援文字的問答數量
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)

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

# 從 DataFrame 中移除欄位
df_view = df_view.drop(columns=columns_to_drop)

# 顯示 DataFrame
df_view.head()

import jsonlines

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

# 寫入 JSONL 檔案
with jsonlines.open('QA_prompts_and_completions.jsonl', 'w') as writer:
    writer.write_all(items)

from openai import OpenAI
import jsonlines

client = OpenAI()

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

# 顯示檔案資訊
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"
)

# 顯示微調任務資訊
print(ft_job)

import pandas as pd
from openai import OpenAI

client = OpenAI()

response = client.fine_tuning.jobs.list(limit=3)

# 初始化要顯示的資訊列表
job_ids = []
created_ats = []
statuses = []
models = []
training_files = []
error_messages = []
fine_tuned_models = []

# 迭代 response 並擷取所需資訊
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_messages.append(job.error)
    fine_tuned_models.append(job.fine_tuned_model)

# 建立 DataFrame
df = pd.DataFrame({
    'Job ID': job_ids,
    'Created At': created_ats,
    'Status': statuses,
    'Model': models,
    'Training File': training_files,
    'Error': error_messages,
    'Fine-Tuned Model': fine_tuned_models
})

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

# 顯示 DataFrame
df

import pandas as pd

generation = False  # 直到目前的模型完成微調

# 篩選出 DataFrame 中 '微調模型' 欄位非空的資料列
non_empty_models = df[df['微調模型'].notna() & (df['微調模型'] != '')]

# 檢查是否有非空的微調模型
if not non_empty_models.empty:
    # 取出第一個非空的微調模型名稱
    first_non_empty_model = non_empty_models['微調模型'].iloc[0]
    print("最新的微調模型是：", first_non_empty_model)
    generation = True
else:
    # 如果沒有非空的微調模型，則將 generation 設為 False
    first_non_empty_model = 'None'
    print("找不到微調模型。")
    generation = False

# 定義提示
prompt = "是什麼現象讓全球風從東北吹向西南？"

# 使用微調模型進行完成任務
if generation:
    response = client.chat.completions.create(
        model=first_non_empty_model,
        temperature=0.0,  # 設定溫度引數，控制輸出的隨機性
        messages=[
            {"role": "system", "content": "給定一個問題，回答問題。"},
            {"role": "user", "content": prompt}
        ]
    )
else:
    print("錯誤：模型為 None，無法繼續 API 請求。")

if generation:
    print(response)

if generation:
    # 從第一個選項中擷取回應文字
    response_text = response.choices[0].message.content
    # 列印回應文字
    print(response_text)

import textwrap

if generation:
    # 將回應文字格式化成段落
    wrapped_text = textwrap.fill(response_text.strip(), 60)
    print(wrapped_text)