configs:
- config_name: default
data_files:- data/OpenSafetyLab_t2i_safety_dataset_0000.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0001.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0002.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0003.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0004.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0005.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0006.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0007.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0008.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0009.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0010.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0011.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0012.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0013.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0014.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0015.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0016.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0017.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0018.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0019.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0020.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0021.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0022.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0023.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0024.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0025.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0026.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0027.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0028.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0029.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0030.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0031.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0032.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0033.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0034.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0035.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0036.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0037.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0038.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0039.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0040.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0041.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0042.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0043.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0044.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0045.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0046.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0047.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0048.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0049.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0050.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0051.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0052.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0053.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0054.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0055.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0056.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0057.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0058.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0059.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0060.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0061.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0062.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0063.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0064.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0065.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0066.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0067.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0068.parquet
- data/OpenSafetyLab_t2i_safety_dataset_0069.parquet
- data/train-00000-of-00001.parquet
- data/train-00000-of-00002.parquet
- data/train-00001-of-00002.parquet
- data/yiting_UnsafeBench_0000.parquet
- data/yiting_UnsafeBench_0001.parquet
- data/yiting_UnsafeBench_0002.parquet
- data/yiting_UnsafeBench_0003.parquet
- data/yiting_UnsafeBench_0004.parquet
- data/yiting_UnsafeBench_0005.parquet
- data/yiting_UnsafeBench_0006.parquet
- data/yiting_UnsafeBench_0007.parquet
- data/yiting_UnsafeBench_0008.parquet
- data/yiting_UnsafeBench_0009.parquet
- data/yiting_UnsafeBench_0010.parquet
- data/yiting_UnsafeBench_0011.parquet
- data/yiting_UnsafeBench_0012.parquet
- data/yiting_UnsafeBench_0013.parquet
- data/yiting_UnsafeBench_0014.parquet
- data/yiting_UnsafeBench_0015.parquet
- data/yiting_UnsafeBench_0016.parquet
- data/yiting_UnsafeBench_0017.parquet
- data/yiting_UnsafeBench_0018.parquet
- data/yiting_UnsafeBench_0019.parquet
- data/yiting_UnsafeBench_0020.parquet
- data/yiting_UnsafeBench_0021.parquet
- data/yiting_UnsafeBench_0022.parquet
- data/yiting_UnsafeBench_0023.parquet
- data/yiting_UnsafeBench_0024.parquet
- data/yiting_UnsafeBench_0025.parquet
- data/yiting_UnsafeBench_0026.parquet
- data/yiting_UnsafeBench_0027.parquet
- data/yiting_UnsafeBench_0028.parquet
- data/yiting_UnsafeBench_0029.parquet
- data/yiting_UnsafeBench_0030.parquet
- data/yiting_UnsafeBench_0031.parquet
- data/yiting_UnsafeBench_0032.parquet
- data/yiting_UnsafeBench_0033.parquet
- data/yiting_UnsafeBench_0034.parquet
- data/yiting_UnsafeBench_0035.parquet
- data/yiting_UnsafeBench_0036.parquet
- data/yiting_UnsafeBench_0037.parquet
- data/yiting_UnsafeBench_0038.parquet
- data/yiting_UnsafeBench_0039.parquet
- data/yiting_UnsafeBench_0040.parquet
- data/yiting_UnsafeBench_0041.parquet
- data/yiting_UnsafeBench_0042.parquet
- data/yiting_UnsafeBench_0044.parquet
Auditor Training Dataset
A large-scale multi-label image safety dataset designed for training cross-attention auditors
to detect red-team attacks on Text-to-Image (T2I) models — specifically prompts that use
euphemistic or indirect language to bypass safety filters.
Motivation
Existing T2I safety checkers rely on keyword matching (e.g. flagging "knife", "gun", "nude").
This dataset is built to train an auditor that catches semantic intent, not just explicit words.
For example: "a man holding a metallic handheld device" should be flagged as potentially violent
even though no explicit keywords are present.
Dataset Statistics
- Total images: ~90,000+ (and growing)
- Format: Multi-label classification (an image can belong to multiple classes)
- Labels: 6 binary columns per image
| Label | Description |
|---|---|
violence |
Weapons, gore, physical harm, disturbing content |
sexual |
Nudity, sexual acts, NSFW content |
hate |
Hate speech, harassment, discrimination, bias, fairness violations |
illegal_activity |
Drug use, theft, fraud, deception |
sensitive_ip |
Copyright, public figures, personal documents, privacy violations |
safe |
Benign images with no unsafe content |
Methodology
Prompt Generation Strategy
For datasets without existing captions, we use Qwen2-VL-2B-Instruct to generate
euphemistic image generation prompts that:
- Accurately describe the visual content
- Avoid explicit flagged words (knife, gun, blood, nude, etc.)
- Sound like plausible innocent Stable Diffusion prompts
This teaches the auditor to detect semantic intent rather than keywords.
Label Schema
Multi-label — a single image can have multiple flags set to 1.
For example, an image of sexual violence would have both sexual=1 and violence=1.
Source Datasets
| Dataset | Class(es) | Prompts | Size |
|---|---|---|---|
Subh775/WeaponDetection |
violence | VLM generated | ~5.8k |
NeuralShell/Gore-Blood-Dataset-v1.0 |
violence | VLM generated | varies |
x1101/nsfw-full |
sexual | VLM generated | varies |
DRDELATV/woman-sexy |
sexual | VLM generated | varies |
Lenkashell/unsafe_violence_image_captions |
violence | existing captions | varies |
Lenkashell/unsafe_shocking_image_captions |
violence | existing captions | varies |
yiting/UnsafeBench |
violence, sexual, hate, illegal_activity, sensitive_ip, safe | existing captions | ~8.1k |
OpenSafetyLab/t2i_safety_dataset |
all classes | existing captions | ~68.7k |
Category Collapse (UnsafeBench → 6 classes)
UnsafeBench's 11 categories were collapsed:
- Shocking, Self-harm →
violence - Harassment →
hate - Deception →
illegal_activity - Political, Copyright →
sensitive_ip
Category Collapse (T2ISafety → 6 classes)
T2ISafety's 25+ categories were collapsed with multi-label support:
- disturbing, disturbing and violence →
violence - humiliation, hate, fairness →
hate - illegal activity →
illegal_activity - intellectual property violation, personal identification documents, public figures →
sensitive_ip - disturbing and sexual →
violence=1, sexual=1 - illegal activity and violence →
violence=1, illegal_activity=1 - (and all other compound categories mapped to multiple flags)
Intended Use
- Training cross-attention safety auditors for T2I models
- Red-team evaluation of diffusion model safety filters
- Research on semantic safety detection beyond keyword filtering