Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -1,34 +1,107 @@
|
|
| 1 |
---
|
| 2 |
license: mit
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 3 |
tags:
|
| 4 |
- url-classification
|
| 5 |
- list-page-detection
|
| 6 |
- detail-page-detection
|
| 7 |
- qwen
|
| 8 |
- fine-tuning
|
| 9 |
-
-
|
|
|
|
| 10 |
widget:
|
| 11 |
- text: "https://example.com/product/12345"
|
|
|
|
| 12 |
---
|
| 13 |
|
| 14 |
# URL Page Type Classifier
|
| 15 |
|
| 16 |
-
|
| 17 |
|
| 18 |
-
|
|
|
|
|
|
|
| 19 |
|
| 20 |
-
|
| 21 |
-
- **微调方法**: LoRA
|
| 22 |
-
- **训练数据**: IowaCat/page_type_inference_dataset (10,000条URL)
|
| 23 |
-
- **训练环境**: NVIDIA RTX 4060 Laptop GPU
|
| 24 |
|
| 25 |
-
##
|
| 26 |
|
| 27 |
-
判断URL是
|
| 28 |
-
- **列表页 (List Page)** - 如 `/products`, `/category`, `/search`
|
| 29 |
-
- **详情页 (Detail Page)** - 如 `/product/12345`, `/item/abc`
|
| 30 |
|
| 31 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
|
| 33 |
```python
|
| 34 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
@@ -37,34 +110,79 @@ model_name = "windlx/url-classifier-model"
|
|
| 37 |
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
| 38 |
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
|
| 39 |
|
|
|
|
| 40 |
url = "https://example.com/product/12345"
|
| 41 |
-
prompt = f"请判断以下URL是列表页还是详情页。\n\nURL: {url}\n类型: "
|
| 42 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 43 |
inputs = tokenizer(prompt, return_tensors="pt")
|
| 44 |
-
outputs = model.generate(**inputs, max_new_tokens=10)
|
| 45 |
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 46 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
```
|
| 48 |
|
| 49 |
-
##
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
|
|
|
| 54 |
|
| 55 |
-
##
|
| 56 |
|
| 57 |
-
-
|
| 58 |
-
-
|
| 59 |
-
-
|
| 60 |
-
- Batch size: 2
|
| 61 |
-
- Learning rate: 2e-4
|
| 62 |
|
| 63 |
-
##
|
| 64 |
|
| 65 |
-
-
|
| 66 |
-
-
|
|
|
|
| 67 |
|
| 68 |
-
## 许可
|
| 69 |
|
| 70 |
-
[LICENSE](LICENSE)
|
|
|
|
| 1 |
---
|
| 2 |
license: mit
|
| 3 |
+
language:
|
| 4 |
+
- zh
|
| 5 |
+
- en
|
| 6 |
+
datasets:
|
| 7 |
+
- IowaCat/page_type_inference_dataset
|
| 8 |
+
metrics:
|
| 9 |
+
- accuracy: 0.99
|
| 10 |
+
pipeline_tag: text-generation
|
| 11 |
tags:
|
| 12 |
- url-classification
|
| 13 |
- list-page-detection
|
| 14 |
- detail-page-detection
|
| 15 |
- qwen
|
| 16 |
- fine-tuning
|
| 17 |
+
- lora
|
| 18 |
+
- url-parser
|
| 19 |
widget:
|
| 20 |
- text: "https://example.com/product/12345"
|
| 21 |
+
- text: "https://example.com/category/electronics"
|
| 22 |
---
|
| 23 |
|
| 24 |
# URL Page Type Classifier
|
| 25 |
|
| 26 |
+
<div align="center">
|
| 27 |
|
| 28 |
+

|
| 29 |
+

|
| 30 |
+

|
| 31 |
|
| 32 |
+
</div>
|
|
|
|
|
|
|
|
|
|
| 33 |
|
| 34 |
+
## 📋 概述
|
| 35 |
|
| 36 |
+
基于 Qwen2.5-1.5B + LoRA 微调的URL类型分类模型,用于判断URL是列表页还是详情页。
|
|
|
|
|
|
|
| 37 |
|
| 38 |
+
## 🏗️ 模型架构
|
| 39 |
+
|
| 40 |
+
| 项目 | 详情 |
|
| 41 |
+
|------|------|
|
| 42 |
+
| **基础模型** | Qwen/Qwen2.5-1.5B |
|
| 43 |
+
| **微调方法** | LoRA (r=16, alpha=32) |
|
| 44 |
+
| **参数量** | 1.5B |
|
| 45 |
+
| **可训练参数** | ~18M (1.18%) |
|
| 46 |
+
|
| 47 |
+
## 📊 训练数据
|
| 48 |
+
|
| 49 |
+
- **数据集**: IowaCat/page_type_inference_dataset
|
| 50 |
+
- **训练样本**: 10,000条URL (5000列表页 + 5000详情页)
|
| 51 |
+
- **数据来源**: HuggingFace Datasets
|
| 52 |
+
|
| 53 |
+
### 数据分布
|
| 54 |
+
|
| 55 |
+
| 类型 | 数量 | 比例 |
|
| 56 |
+
|------|------|------|
|
| 57 |
+
| 列表页 (List Page) | 5,000 | 50% |
|
| 58 |
+
| 详情页 (Detail Page) | 5,000 | 50% |
|
| 59 |
+
|
| 60 |
+
## ⚙️ 训练配置
|
| 61 |
+
|
| 62 |
+
```python
|
| 63 |
+
{
|
| 64 |
+
"base_model": "Qwen/Qwen2.5-1.5B",
|
| 65 |
+
"lora_rank": 16,
|
| 66 |
+
"lora_alpha": 32,
|
| 67 |
+
"lora_dropout": 0.05,
|
| 68 |
+
"num_train_epochs": 3,
|
| 69 |
+
"per_device_train_batch_size": 2,
|
| 70 |
+
"gradient_accumulation_steps": 8,
|
| 71 |
+
"learning_rate": 2e-4,
|
| 72 |
+
"fp16": true,
|
| 73 |
+
"optimizer": "adamw_torch",
|
| 74 |
+
"lr_scheduler_type": "cosine"
|
| 75 |
+
}
|
| 76 |
+
```
|
| 77 |
+
|
| 78 |
+
## 📈 性能评估
|
| 79 |
+
|
| 80 |
+
### 测试结果
|
| 81 |
+
|
| 82 |
+
| 测试集 | 样本数 | 准确率 |
|
| 83 |
+
|--------|--------|--------|
|
| 84 |
+
| 验证集 | 100 | **99%** |
|
| 85 |
+
|
| 86 |
+
### 示例预测
|
| 87 |
+
|
| 88 |
+
| URL | 预测结果 |
|
| 89 |
+
|-----|----------|
|
| 90 |
+
| `https://example.com/products/category` | 列表页 (List Page) |
|
| 91 |
+
| `https://example.com/product/12345` | 详情页 (Detail Page) |
|
| 92 |
+
| `https://example.com/search?q=test` | 列表页 (List Page) |
|
| 93 |
+
| `https://example.com/item/abc123` | 详情页 (Detail Page) |
|
| 94 |
+
| `https://example.com/list/all` | 列表页 (List Page) |
|
| 95 |
+
|
| 96 |
+
## 🚀 快速开始
|
| 97 |
+
|
| 98 |
+
### 安装依赖
|
| 99 |
+
|
| 100 |
+
```bash
|
| 101 |
+
pip install transformers peft torch
|
| 102 |
+
```
|
| 103 |
+
|
| 104 |
+
### 推理代码
|
| 105 |
|
| 106 |
```python
|
| 107 |
from transformers import AutoTokenizer, AutoModelForCausalLM
|
|
|
|
| 110 |
tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
|
| 111 |
model = AutoModelForCausalLM.from_pretrained(model_name, trust_remote_code=True)
|
| 112 |
|
| 113 |
+
# 要分类的URL
|
| 114 |
url = "https://example.com/product/12345"
|
|
|
|
| 115 |
|
| 116 |
+
# 构建提示
|
| 117 |
+
prompt = f"""请判断以下URL是列表页还是详情页。
|
| 118 |
+
|
| 119 |
+
URL: {url}
|
| 120 |
+
类型: """
|
| 121 |
+
|
| 122 |
+
# 推理
|
| 123 |
inputs = tokenizer(prompt, return_tensors="pt")
|
| 124 |
+
outputs = model.generate(**inputs, max_new_tokens=10, do_sample=False)
|
| 125 |
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
|
| 126 |
+
|
| 127 |
+
# 提取结果
|
| 128 |
+
if "详情页" in response or "Detail Page" in response:
|
| 129 |
+
result = "详情页 (Detail Page)"
|
| 130 |
+
else:
|
| 131 |
+
result = "列表页 (List Page)"
|
| 132 |
+
|
| 133 |
+
print(f"URL: {url}")
|
| 134 |
+
print(f"类型: {result}")
|
| 135 |
+
```
|
| 136 |
+
|
| 137 |
+
### 使用 GPU
|
| 138 |
+
|
| 139 |
+
```python
|
| 140 |
+
# 自动使用GPU
|
| 141 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 142 |
+
model_name,
|
| 143 |
+
trust_remote_code=True,
|
| 144 |
+
device_map="auto",
|
| 145 |
+
torch_dtype="auto"
|
| 146 |
+
)
|
| 147 |
+
```
|
| 148 |
+
|
| 149 |
+
### 使用 CPU
|
| 150 |
+
|
| 151 |
+
```python
|
| 152 |
+
# 强制使用CPU
|
| 153 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 154 |
+
model_name,
|
| 155 |
+
trust_remote_code=True,
|
| 156 |
+
device_map="cpu",
|
| 157 |
+
torch_dtype="float32"
|
| 158 |
+
)
|
| 159 |
```
|
| 160 |
|
| 161 |
+
## ⚠️ 局限性
|
| 162 |
+
|
| 163 |
+
1. **仅基于URL字符串** - 不访问实际网页内容
|
| 164 |
+
2. **依赖URL路径规范** - 对于URL路径不规范的网站,准确率可能较低
|
| 165 |
+
3. **仅支持中英文** - 主要针对中文URL优化
|
| 166 |
+
|
| 167 |
+
## 📝 使用场景
|
| 168 |
|
| 169 |
+
- 🔍 **搜索引擎优化 (SEO)** - 识别网站页面结构
|
| 170 |
+
- 🕷️ **网页爬虫** - 判断链接类型,优化爬取策略
|
| 171 |
+
- 📊 **网站分析** - 统计列表页和详情页比例
|
| 172 |
+
- 🔗 **链接分类** - 大规模URL分类处理
|
| 173 |
|
| 174 |
+
## 📁 相关链接
|
| 175 |
|
| 176 |
+
- **GitHub仓库**: https://github.com/xiuxiu/url-classifier
|
| 177 |
+
- **HuggingFace模型**: https://huggingface.co/windlx/url-classifier-model
|
| 178 |
+
- **训练数据集**: https://huggingface.co/datasets/IowaCat/page_type_inference_dataset
|
|
|
|
|
|
|
| 179 |
|
| 180 |
+
## 🙏 致谢
|
| 181 |
|
| 182 |
+
- [Qwen](https://github.com/QwenLM/Qwen2) - 提供基础模型
|
| 183 |
+
- [LoRA](https://github.com/microsoft/LoRA) - 高效微调方法
|
| 184 |
+
- [HuggingFace](https://huggingface.co/) - 模型托管平台
|
| 185 |
|
| 186 |
+
## 📄 许可
|
| 187 |
|
| 188 |
+
[LICENSE](LICENSE) - MIT License
|