README.md

---
license: mit
language:
  - zh
task_categories:
  - text-classification
tags:
  - safety
  - adversarial
  - traditional-chinese
  - content-moderation
  - llm-safety
pretty_name: PATCH (Prompt Assortment for Traditional Chinese Hazards)
size_categories:
  - 100K<n<1M
---

# PATCH: Prompt Assortment for Traditional Chinese Hazards

The first large-scale adversarial safety dataset for Traditional Chinese (TC), designed to train and evaluate content safety classifiers for lightweight LLMs.

For full documentation, see our [GitHub repository](https://github.com/Harrychangtw/PATCH).

## Dataset Overview

- **593,020 safe prompts** localized to Traditional Chinese
- **231,924 unsafe prompts** across 13 [MLCommons](https://arxiv.org/abs/2404.12241) hazard categories
  - **PATCH-GPT**: Direct harmful prompts
  - **PATCH-RT**: Evasive prompts exploiting TC-specific cultural/linguistic patterns
- **PATCH-H**: Gold-standard human-annotated benchmark (390 prompts, Fleiss' κ = 0.84) — held out from training

## Dataset Structure

```
├── safe/           # PATCH_safe_{train,val,test}.csv
├── unsafe_gpt/     # PATCH_unsafe_gpt_{train,val,test}.csv
└── unsafe_rt/      # PATCH_unsafe_rt_{train,val,test}.csv
```

All files follow a 70:10:20 train/val/test split.

## Usage

```python
from datasets import load_dataset

dataset = load_dataset("Raymond102103028/PATCH")
```

## License

[MIT License](https://opensource.org/licenses/MIT) — for research use only.