LLM合成数据生成的全面指南

from transformers import AutoTokenizer, AutoModelForCausalLM

<p># 加载预训练的 LLM
model_name = "gpt2-large"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)</p>

<p># 定义合成数据生成的提示
prompt = "生成一篇关于智能手机的客户评论：" </p>

<p># 生成合成数据
input_ids = tokenizer.encode(prompt, return_tensors="pt")
output = model.generate(input_ids, max_length=100, num_return_sequences=1)</p>

<p># 解码和打印生成的文本
synthetic_review = tokenizer.decode(output[0], skip_special_tokens=True)
print(synthetic_review)

prompt = """
生成一篇关于智能手机的详细客户评论，具有以下特征：
- 品牌：{brand}
- 型号：{model}
- 关键特征：{features}
- 评分：{rating}/5 星

<p>评论应在 50-100 个字之间，包括正面和负面评价。</p>

评论：
"""
brands = ["Apple", "Samsung", "Google", "OnePlus"]
models = ["iPhone 13 Pro", "Galaxy S21", "Pixel 6", "9 Pro"]
features = ["5G, OLED 显示屏, 三摄像头", "120Hz 刷新率, 8K 视频", "AI 驱动的相机, 5G", "快速充电, 120Hz 显示屏"]
ratings = [4, 3, 5, 4]

<p># 生成多个评论
for brand, model, feature, rating in zip(brands, models, features, ratings):
filled_prompt = prompt.format(brand=brand, model=model, features=feature, rating=rating)
input_ids = tokenizer.encode(filled_prompt, return_tensors="pt")
output = model.generate(input_ids, max_length=200, num_return_sequences=1)
synthetic_review = tokenizer.decode(output[0], skip_special_tokens=True)
print(f"评论 for {brand} {model}:\n{synthetic_review}\n")

few_shot_prompt = """
生成一段关于产品问题的客户支持对话，格式如下：

<p>C：您好，我正在使用新耳机遇到问题。右耳的耳塞不工作。
A：抱歉听到这个消息。您能告诉我您使用的耳机型号吗？
C：它是 SoundMax Pro 3000。
A：谢谢。您是否尝试过将耳机放入充电盒 10 秒钟来重置？
C：是的，我已经尝试过了，但没有帮助。
A：我明白了。让我们尝试更新固件。您能访问我们的网站并下载最新的固件吗？

<p>现在生成一段关于不同产品问题的新对话：</p>

<p>C：嗨，我刚刚收到新智能手表，但它无法开机。
"""

from transformers import GPT2LMHeadModel, GPT2Tokenizer
import torch

<p>model = GPT2LMHeadModel.from_pretrained("gpt2-medium")
tokenizer = GPT2Tokenizer.from_pretrained("gpt2-medium")</p>

<p>def generate_conditional_text(prompt, condition, max_length=100):
input_ids = tokenizer.encode(prompt, return_tensors="pt")
attention_mask = torch.ones(input_ids.shape, dtype=torch.long, device=input_ids.device)</p>

<p># 编码条件
condition_ids = tokenizer.encode(condition, add_special_tokens=False, return_tensors="pt")</p>

<p># 连接条件和输入 ID
input_ids = torch.cat([condition_ids, input_ids], dim=-1)
attention_mask = torch.cat([torch.ones(condition_ids.shape, dtype=torch.long, device=condition_ids.device), attention_mask], dim=-1)</p>

<p>output = model.generate(input_ids, attention_mask=attention_mask, max_length=max_length, num_return_sequences=1, no_repeat_ngram_size=2, do_sample=True, top_k=50, top_p=0.95, temperature=0.7)</p>

<p>return tokenizer.decode(output[0], skip_special_tokens=True)</p>

<p># 生成具有不同条件的产品描述
conditions = ["豪华", "经济实惠", "环保", "高科技"]
prompt = "描述一个背包：" </p>

<p>for condition in conditions:
description = generate_conditional_text(prompt, condition)
print(f"{condition} 背包描述：\n{description}\n")

import pandas as pd
from sklearn.model_selection import train_test_split
from transformers import pipeline

<p># 加载一个小型现实世界数据集
real_data = pd.read_csv("small_product_reviews.csv")</p>

<p># 分割数据
train_data, test_data = train_test_split(real_data, test_size=0.2, random_state=42)</p>

<p># 初始化文本生成管道
generator = pipeline("text-generation", model="gpt2-medium")</p>

<p>def augment_dataset(data, num_synthetic_samples):
synthetic_data = []
for _, row in data.iterrows():
prompt = f"生成一篇类似于 {row['review']} 的产品评论\n新评论："
synthetic_review = generator(prompt, max_length=100, num_return_sequences=1)[0]['generated_text']
synthetic_data.append({'review': synthetic_review, 'sentiment': row['sentiment'] # 假设情绪保持不变
if len(synthetic_data) &gt;= num_synthetic_samples:
break
return pd.DataFrame(synthetic_data)</p>

<p># 生成合成数据
synthetic_train_data = augment_dataset(train_data, num_synthetic_samples=len(train_data))</p>

<p># 合并真实和合成数据
augmented_train_data = pd.concat([train_data, synthetic_train_data], ignore_index=True)</p>

<p>print(f"原始训练数据大小：{len(train_data)}")
print(f"增强训练数据大小：{len(augmented_train_data)}")</p>