configs:
- config_name: overt-full
data_files:- split: default
path: data/OVERT_full.csv
- split: default
- config_name: overt-mini
data_files:- split: default
path: data/OVERT_mini.csv
- split: default
- config_name: overt-unsafe
data_files:- split: default
path: data/OVERT_unsafe.csv
- split: default
OVERT: Over-Refusal Evaluation on Text-to-Image Models
Table of Contents
Introduction
We introduce OVERT (OVEr-Refusal evaluation on Text-to-image models), the first large-scale benchmark to evaluate safety over-refusal in T2I models. OVERT includes 4,600 seemingly harmful but benign prompts across nine safety-related categories, along with 1,785 genuinely harmful prompts (OVERT-unsafe) to evaluate the safety–utility trade-off. Using OVERT, we evaluate several leading T2I models and find that over-refusal is a widespread issue across various categories (Figure 3), underscoring the need for further research to enhance the safety alignment of T2I models without compromising their functionality. You can find a summary of the dataset below:
Figure 1: Left: Category distribution of the 4,600 prompts in OVERT. Right: A benign prompt from OVERT is refused by FLUX1.1-Pro and DALL-E-3, but accepted by Imagen-3 and SD-3.5.
Overall workflow
We present an automatic workflow to construct OVERT. You can check out the data under directory data.
Figure 2: OVERT dataset construction pipeline. Prompts are generated via LLMs from WildGuardMix or templates, filtered and audited for safety, deduplicated, and sampled using Chameleon. The final dataset is used to evaluate over-refusal in T2I models.
Evaluation results
Figure 3: Refusal rates of Text-to-Image (T2I) models on benign prompts (x-axis, OVERT-mini) and safe response rate on harmful prompts (y-axis, OVERT-unsafe), grouped into four broad safety categories. Each point corresponds to a specific model's refusal rate within one broad category, obtained by aggregating across related subsets of the nine fine-grained categories. The dashed curve shows a quadratic regression fit, highlighting the trade-off between safety and over-refusal.
Figure 4: Refusal rates (percentage of requests refused by the models) of five T2I models across nine categories on (a) OVERT-mini (benign prompts) and (b) OVERT-unsafe (harmful prompts). The results show that different models exhibit distinct refusal behaviors across categories.
You can reproduce the evaluation results by codes in our GitHub repo