更新数据集结构: 添加parquet格式和重新组织的图像文件

README.md

@@ -20,137 +20,126 @@ tags:
   - xiaohongshu
   - RedNote
 pretty_name: CHASM - Covert Advertisement on RedNote
+configs:
+  - config_name: default
+    default: true
+    data_files:
+      - split: train
+        path: "parquet_data_optimized/train.parquet"
+      - split: validation
+        path: "parquet_data_optimized/validation.parquet"
+      - split: test
+        path: "parquet_data_optimized/test.parquet"
 ---
 
 <!-- @format -->
 
-# CHASM: Covert Advertisement on RedNote Dataset
+# CHASM Dataset
 
-A dataset containing posts from Xiaohongshu (RedNote) for text classification tasks, specifically focused on identifying covert advertisements.
+CHASM (Covert Advertisement on RedNote) is a dataset containing RedNote posts for text and image classification tasks focused on identifying covert advertisements. The dataset includes samples categorized as advertisements (label=1) and non-advertisements (label=0).
 
-## Dataset Description
+## Dataset Format
 
-- **Task Type**: Text Classification (Binary Classification)
-- **Language**: Chinese
-- **License**: MIT
-- **Dataset Size**: Contains two classes (label_0 and label_1)
-- **Splits**: Train (3089 samples), Validation (440 samples), Test (888 samples)
+The dataset is provided in an optimized parquet format:
 
-### Dataset Structure
-
-This dataset is available in two formats:
-
-#### 1. Original Directory Structure
-
-Data is organized in the following directory structure:
-
-- `label_0/`: Non-advertisement content
-- `label_1/`: Advertisement content
-
-Each sample has its own folder (e.g., `train_889/`), containing:
-
-- `data.json`: Text data
-- Multiple image files (JPG/WEBP format)
+```
+parquet_data_optimized/
+├── train.parquet       # Training set
+├── validation.parquet  # Validation set
+└── test.parquet        # Test set
+```
 
-#### 2. Hugging Face Format
+Images are stored in a structured directory system:
 
-For better compatibility with the Hugging Face platform, we also provide processed formats:
+```
+processed_images/
+├── images_0000/        # Image subfolder 0
+├── images_0001/        # Image subfolder 1
+└── ...                 # Additional image subfolders
+```
 
-- `hf_format/train.csv` and `hf_format/train.json`: Training set
-- `hf_format/validation.csv` and `hf_format/validation.json`: Validation set
-- `hf_format/test.csv` and `hf_format/test.json`: Test set
-- `hf_format/dataset_info.json`: Dataset metadata
+Each image is named following the pattern `post_{id}_img_{index}.jpg`, where:
 
-### Data Fields
+- `id` is the unique numeric identifier of the post
+- `index` is the image index within that post
 
-Each data sample contains the following fields:
+The parquet files contain the following fields:
 
-- `id`: Sample unique identifier
+- `id`: Numeric sample identifier (sequentially assigned)
 - `title`: Post title
-- `description`: Post description
-- `date`: Publication date and location
-- `comments`: List of comments
-- `images`: List of image filenames
+- `description`: Post description content
+- `date`: Published date and location
+- `comments`: List of comments (as array)
+- `images`: List of image paths relative to processed_images directory
+- `image_count`: Number of images in the post
 - `label`: Classification label (0: non-advertisement, 1: advertisement)
+- `split`: Dataset split (train, validation, test)
 
-### Data Example
-
-```json
-{
-    "id": "train_889",
-    "title": "豆瓣9.2❗️几乎是全程震撼着读完的一本书❗️",
-    "description": "📚《人类新史》\n📝［美］大卫·格雷伯/［英］大卫·温格罗\n\n说实话我几乎是不读社科历史大部头的，一年能读上一本我都想夸自己厉害🤣所以看到这本书的厚度时心里真的发怵，担心自己读不下去，却没想到它出乎意料地好读，才读了百来页我就忍不住先来分享一波了，因为脑袋里信息量爆表了！！...",
-    "date": "编辑于 7 天前 湖南",
-    "comments": ["感谢推荐，已收藏", "你好厉害，能看英文版我看中文版才看了这么点就觉得很费劲", ...],
-    "images": ["1040g008317prti2fk45g48iu7h5portohqvdpdo!nd_dft_wlteh_webp_3_0.jpg", ...],
-    "label": 0
-}
-```
+## Dataset Statistics
+
+- **Total samples**: 4,992 samples
+- **Class distribution**:
+  - Non-advertisements (label=0): 4,379 samples (87.7%)
+  - Advertisements (label=1): 613 samples (12.3%)
+- **Splits**:
+  - Training set: 3,493 samples
+  - Validation set: 499 samples
+  - Test set: 1,000 samples
+- **Images**: 26,324 images in total
+- **Average images per post**: 5.27 images
 
 ## Usage
 
-Using the original directory structure:
+### Loading Parquet Data
 
 ```python
-import os
-import json
-from glob import glob
+import pandas as pd
 
-# Iterate through all samples
-for label_dir in ['label_0', 'label_1']:
-    for sample_dir in glob(f"{label_dir}/*"):
-        if os.path.isdir(sample_dir):
-            # Read data file
-            with open(os.path.join(sample_dir, 'data.json'), 'r', encoding='utf-8') as f:
-                data = json.load(f)
+# Load training set
+train_df = pd.read_parquet("parquet_data_optimized/train.parquet")
 
-            # Get image files
-            images = glob(os.path.join(sample_dir, '*.jpg'))
+# View data
+print(train_df.head())
+print(f"Training samples: {len(train_df)}")
 
-            # Process data...
+# Access images
+print(f"First sample images: {train_df.iloc[0]['images']}")
+print(f"Image count: {train_df.iloc[0]['image_count']}")
 ```
 
-Using Hugging Face format:
+### Accessing Images
 
 ```python
-import pandas as pd
-
-# Load data
-train_df = pd.read_csv('hf_format/train.csv')
-val_df = pd.read_csv('hf_format/validation.csv')
-test_df = pd.read_csv('hf_format/test.csv')
-
-# Or use JSON format
-import json
-with open('hf_format/train.json', 'r', encoding='utf-8') as f:
-    train_data = json.load(f)
+import os
+from PIL import Image
+import matplotlib.pyplot as plt
+
+# Function to load and display an image
+def load_and_display_image(image_path, base_dir="processed_images"):
+    full_path = os.path.join(base_dir, image_path)
+    img = Image.open(full_path)
+    plt.figure(figsize=(10, 8))
+    plt.imshow(img)
+    plt.axis('off')
+    plt.title(f"Image: {image_path}")
+    plt.show()
+
+# Example: load the first image of the first post in the training set
+if len(train_df) > 0 and len(train_df.iloc[0]['images']) > 0:
+    first_image_path = train_df.iloc[0]['images'][0]
+    load_and_display_image(first_image_path)
 ```
 
-Using Hugging Face datasets library:
+## Citation
 
-```python
-from datasets import load_dataset
-
-# Load dataset
-dataset = load_dataset("Jingyi77/CHASM-Covert_Advertisement_on_RedNote")
+If you use the CHASM dataset, please cite:
 
-# View dataset information
-print(dataset)
-
-# Access data
-print(dataset['train'][0])  # View first data point in training set
 ```
-
-## Data Preprocessing
-
-- Image paths have been simplified to filenames only
-- All sensitive information (phone numbers, ID numbers, etc.) has been anonymized
-- Dates and locations have been partially anonymized
-
-## Notes
-
-This dataset is specifically designed for research on covert advertisement detection as a binary classification task.
-
-## Dataset Creation
-
-The dataset was created by collecting and processing posts from Xiaohongshu (RedNote). Each post was manually labeled to identify covert advertisements. The dataset has been carefully anonymized to protect user privacy while maintaining the essential characteristics needed for research purposes.
+@misc{CHASM2023,
+  author = {Jingyi},
+  title = {CHASM: Covert Advertisement on RedNote},
+  year = {2023},
+  publisher = {Hugging Face},
+  howpublished = {\url{https://huggingface.co/datasets/Jingyi77/CHASM-Covert_Advertisement_on_RedNote}}
+}
+```