Highgroundbkk/ThaiSafetyBench

---
license: apache-2.0
viewer: false
task_categories:
- question-answering
language:
- th
size_categories:
- 1K<n<10K
dataset_info:
features:
- name: prompt
dtype: string
- name: risk_area
dtype: string
- name: types_of_harm
dtype: string
- name: subtypes_of_harm
dtype: string
- name: thai_related
dtype: bool
- name: ai_gen
dtype: bool
- name: source
dtype: string
- name: id
dtype: int64
splits:
- name: test
num_bytes: 549743
num_examples: 1889
download_size: 137554
dataset_size: 565298
configs:
- config_name: default
data_files:
- split: test
path: data/test-*
---

# ThaiSafetyBench

<span style="color: red;">⚠️ **Warning:** This dataset contains harmful and toxic language. It is intended for academic purposes only.</span>

[[ArXiv Paper](https://arxiv.org/abs/2603.04992v2)] [[Github](https://github.com/trapoom555/ThaiSafetyBench)] [[Hugging Face Leaderboard 🤗](https://huggingface.co/spaces/typhoon-ai/ThaiSafetyBench-Leaderboard)]

The ThaiSafetyBench dataset comprises 1,889 malicious Thai-language prompts across various categories. In addition to translated malicious prompts, it includes prompts tailored to Thai culture, offering deeper insights into culturally specific attacks.

> **Note:** The *Monarchy* type of harm has been removed from the dataset to comply with Thai regulations. This filtering reduced the dataset from 1,954 to 1,889 samples.

## Dataset Distribution

The dataset is categorized into hierarchical categories, as illustrated in the figure below.


## Dataset Fields

| **Field** | **Description** |
|--------------------|---------------------------------------------------------------------------------|
| `id` | A unique identifier for each sample in the dataset. |
| `prompt` | The malicious prompt written in Thai. |
| `risk_area` | The primary hierarchical category of risk associated with the prompt. |
| `types_of_harm` | A subcategory of the `risk_area`, specifying the type of harm. |
| `subtypes_of_harm` | A further subcategory under `types_of_harm`, providing more granular detail. |
| `thai_related` | A boolean (`true`/`false`) indicating whether the prompt is related to Thai culture. |
| `ai_gen` | A boolean (`true`/`false`) indicating whether the prompt was generated by AI. |
| `source` | The source from which the data was collected. |

**Note:** The hierarchical structure (`risk_area` → `types_of_harm` → `subtypes_of_harm`) allows for detailed classification of the potential harm caused by each prompt.

## More Dataset Statistics


## Citation

```
@misc{ukarapol2026thaisafetybenchassessinglanguagemodel,
title={ThaiSafetyBench: Assessing Language Model Safety in Thai Cultural Contexts},
author={Trapoom Ukarapol and Nut Chukamphaeng and Kunat Pipatanakul and Pakhapoom Sarapat},
year={2026},
eprint={2603.04992},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2603.04992},
}
```