README.md

---
language:
- zh
license: mit
size_categories:
- 1K<n<10K
pretty_name: CHASM - RedNote Covert Advertisement Detection Dataset
task_categories:
- image-text-to-text
tags:
- covert advertisement detection
- social-media
- image-text
- multimodal
- RedNote
- Xiaohongshu
datasets:
- Jingyi77/CHASM-Covert_Advertisement_on_RedNote
dataset_info:
- config_name: default
  features:
  - name: id
    dtype: string
  - name: title
    dtype: string
  - name: description
    dtype: string
  - name: date
    dtype: string
  - name: comments
    sequence:
      dtype: string
  - name: images
    sequence:
      dtype: string
  - name: image_count
    dtype: int32
  - name: label
    dtype: int8
  - name: split
    dtype: string
configs:
- config_name: Example
  data_files:
  - split: Example
    path: example.parquet
  - split: Train_1
    path: train_part_1.parquet
  - split: Train_2
    path: train_part_2.parquet
  - split: Train_3
    path: train_part_3.parquet
  - split: Train_4
    path: train_part_4.parquet
  - split: Test
    path: test.parquet
  - split: Validation
    path: validation.parquet
---

<!-- @format -->

# RedNote Covert Advertisement Detection Dataset (CHASM)

This dataset contains posts from the RedNote platform for covert advertisement detection tasks, introduced in the paper [CHASM: Unveiling Covert Advertisements on Chinese Social Media](https://huggingface.co/papers/2604.20511).

[**Paper**](https://huggingface.co/papers/2604.20511) | [**Code**](https://github.com/Jingyi62/CHASM)

## Dataset Overview

| Split      | Posts    | Ad Posts | Non-Ad Posts | Total Images |
| ---------- | -------- | -------- | ------------ | ------------ |
| Train      | 3493     | 426      | 3067         | 18543        |
| Validation | 499      | 57       | 442          | 2678         |
| Test       | 1000     | 130      | 870          | 5103         |
| **Total**  | **4992** | **613**  | **4379**     | **26324**    |

> Note: The viewer shows a **small example subset** of the data (60 samples) for demonstration purposes. The complete dataset is available via WebDataset format in the repository.

## Field Descriptions

The example parquet file contains the following fields:

- `id`: Unique identifier for each post
- `title`: Post title
- `description`: Post description content
- `date`: Publication date (format: MM-DD)
- `comments`: List of comments
- `images`: List of base64-encoded images
- `image_count`: Number of images
- `label`: Label (0=non-advertisement, 1=advertisement)
- `split`: Data split (train/validation/test)

## Dataset Features

- **Multimodal Data**: Each post contains both text (title, description, comments) and images.
- **Real-world Data**: Collected from actual social media posts on the RedNote platform.
- **Multiple Images**: Each post may contain multiple images (average of 5.27 images per post).

## Data Format

The complete dataset is stored in WebDataset format, with each sample containing:

1. One or more image files (.jpg format)
2. A JSON metadata file with the following fields:
   - `id`: Sample ID
   - `title`: Title
   - `description`: Description
   - `date`: Date
   - `comments`: List of comments
   - `label`: Label (0: non-advertisement, 1: advertisement)

## Loading the Dataset

```python
from datasets import load_dataset

# Load example dataset
dataset = load_dataset("Jingyi77/CHASM-Covert_Advertisement_on_RedNote")

# Access a sample
example = dataset[0]
metadata = {
    "id": example["id"],
    "title": example["title"],
    "description": example["description"],
    "label": example["label"]
}
images = example["images"]  # List of images
```

## Citation

If you use this dataset in your research, please cite:

```bibtex
@article{zheng2025chasm,
  title={CHASM: Unveiling Covert Advertisements on Chinese Social Media},
  author={Zheng, Jingyi and Hu, Tianyi and Liu, Yule and Sun, Zhen and Zhang, Zongmin and Dong, Wenhan and Peng, Zifan and He, Xinlei},
  journal={arXiv preprint arXiv:2604.20511},
  year={2025}
}
```