Unsafe-1K
Unsafe-1K is an evaluation dataset containing 1,039 unsafe text prompts for studying the safety of text-to-image generation models. The dataset is designed to provide a controlled yet diverse benchmark for evaluating whether image generation systems can recognize and handle unsafe requests.
Dataset Construction
Unsafe-1K is constructed using a modifier-based prompt composition strategy adapted from MODX.
Each prompt is generated from three components:
- Unsafe subjects: sensitive or unsafe concepts collected from public datasets and manually curated seed terms.
- Modifiers: style, medium, and descriptive expressions that can affect how unsafe content is represented.
- Prompt templates: diverse sentence structures used to reduce linguistic bias while preserving the original semantics.
The prompt construction process can be summarized as:
Prompt = Template(Unsafe Subject, Modifier)