---
task_categories:
- text-classification
language:
- vi
---
## Dataset Card for VITHSD

### 1. Dataset Summary

**VITHSD** (Vietnamese Targeted Hate Speech Detection) contains **10,000** Vietnamese social‐media comments annotated for hate toward five target categories:

* **individual**
* **groups**
* **religion/creed**
* **race/ethnicity**
* **politics**

Each target is labeled on a **3‐point scale** (e.g., 0 = no hate, 1 = offensive, 2 = hateful). In this unified version, all splits are combined into one CSV with an extra `type` column indicating `train` / `dev` / `test`.


### 2. Supported Tasks and Metrics

* **Task**: Multi‐label classification (per‐target hate/offense detection)
* **Metrics**:

  * **Accuracy** (per‐target and overall)
  * **Macro F1** (per‐target and overall)


### 3. Languages

* Vietnamese

### 4. Dataset Structure

| Column           | Type   | Description                                  |
| ---------------- | ------ | -------------------------------------------- |
| `comment`        | string | The raw user comment (Vietnamese).           |
| `individual`     | int    | Hate level toward individuals (0/1/2).       |
| `groups`         | int    | Hate level toward groups (0/1/2).            |
| `religion/creed` | int    | Hate level toward religion/creed (0/1/2).    |
| `race/ethnicity` | int    | Hate level toward race/ethnicity (0/1/2).    |
| `politics`       | int    | Hate level toward political targets (0/1/2). |
| `type`           | string | Split name: `train` / `validation` / `test`. |
| `dataset`        | string | Source indicator (`VITHSD`).                 |


### 5. Data Fields

* **comment** (`str`): The social‐media text.
* **individual**, **groups**, **religion/creed**, **race/ethnicity**, **politics** (`int`): 0 = no hate, 1 = offensive, 2 = hateful.
* **type** (`str`): Which split the example belongs to.
* **dataset** (`str`): Always `VITHSD`.

### 6. Usage

```python
from datasets import load_dataset

ds = load_dataset("visolex/VITHSD")

train = ds.filter(lambda ex: ex["type"] == "train")
val   = ds.filter(lambda ex: ex["type"] == "dev")
test  = ds.filter(lambda ex: ex["type"] == "test")

# Examine one example
print(train[0])
```


### 7. Source & Links

* **Hugging Face Dataset**
  [`sonlam1102/vithsd`](https://huggingface.co/datasets/sonlam1102/vithsd)
* **Publication / Preprint**
  Vo et al. (2024), “ViTHSD: exploiting hatred by targets for hate speech detection on Vietnamese social media texts” (arXiv)

### 8. Contact Information

* **Mr. Son T. Luu**
* Email: [sonlt@uit.edu.vn](mailto:sonlt@uit.edu.vn)
* Alternative: [son.lt1103@gmail.com](mailto:son.lt1103@gmail.com)


### 9. Licensing and Citation

#### License

Refer to the original repository’s LICENSE. If unspecified, assume **CC BY 4.0**.

#### How to Cite

```bibtex
@article{vo2025vithsd,
  title   = {ViTHSD: exploiting hatred by targets for hate speech detection on Vietnamese social media texts},
  author  = {Vo, Cuong Nhat and Huynh, Khanh Bao and Luu, Son T and Do, Trong-Hop},
  journal = {Journal of Computational Social Science},
  volume  = {8},
  number  = {2},
  pages   = {30},
  year    = {2025},
  publisher = {Springer},
  doi     = {10.1007/s42001-024-00348-6}
}
```

```bibtex
@misc{sonlam1102_vithsd,
  title        = {VITHSD: Targeted Hate Speech Detection on Vietnamese Social Media Texts},
  howpublished = {\url{https://huggingface.co/datasets/sonlam1102/vithsd}},
  year         = {2024}
}
```