license: mit
pretty_name: RW2Unsafe Prompts
language:
- en
size_categories:
- 1K<n<10K
extra_gated_prompt: >-
To request access, provide your ORCID iD and email address. Requests are
reviewed by the dataset authors before access is granted.
extra_gated_fields:
ORCID iD: text
Email address: text
tags:
- safety
- text-to-image
- generated
configs:
- config_name: default
data_files:
- split: benchmark
path: rewritten_prompts_and_intents.json
rw_t2i_4chan
Content warning: This dataset contains rewritten prompts describing hateful, harassing, violent, sexually explicit, self-harm-related, and other unsafe content. Reading it may be disturbing.
Dataset description
rw_t2i_4chan is a collection of 2,000 rewritten text-to-image prompts for safety evaluation, introduced in RW2Unsafe: Towards More Accurate Real-World Unsafe Generation Evaluation for T2I Models. The prompts are derived from publicly available 4chan posts and rewritten to describe visual scenes. The intended use is research on the safety of text-to-image systems.
Fields and distribution
Each record has two string fields:
prompt: rewritten text-to-image prompt.malicious_intent: assigned harmful-intent category.
| Label | Records |
|---|---|
| sexually explicit content | 865 |
| hate speech | 675 |
| harassment | 298 |
| violent content | 90 |
| dangerous and illegal activities | 44 |
| self-harm | 28 |
The dataset is provided as one benchmark split containing all 2,000 records; it has not been divided into separate training, validation, or test sets.
Use
To request access, submit your ORCID iD and email address through the Hugging Face access request form. The dataset authors review requests manually before granting access. Contact the first author listed in the paper for questions. Request information is used to review access and communicate about the request. Handle the material with care and follow applicable institutional and platform policies.
Loading
After access is granted, load the dataset with:
from datasets import load_dataset
dataset = load_dataset("TrustAIRLab/RW2Unsafe_dataset", split="benchmark")
print(len(dataset), dataset.column_names)
License and citation
The repository metadata specifies the MIT license. Users must also respect any applicable third-party rights and platform terms.
Ye Leng, Junjie Chu, Mingjie Li, Michael Backes, Yun Shen, and Yang Zhang. 2026. RW2Unsafe: Towards More Accurate Real-World Unsafe Generation Evaluation for T2I Models. Findings of the Association for Computational Linguistics: EMNLP 2026, pages 28860–28873.