SafeAtlas-VL contains multimodal content spanning a range of safety risks. By requesting access, you agree to use the dataset only for lawful and responsible research, evaluation, or educational purposes.

SafeAtlas-VL

SafeAtlas-VL is a large-scale image-grounded safety dataset with five-level ordinal safety labels for image, request, and response moderation.

This release provides a 1,000,000-annotation training set covering 594,487 unique images. The complete SafeAtlas-VL dataset will be released in a future update.

Dataset size

Split	Unique images	Safety annotations
train	594,487	1,000,000
test	4,979	5,000

Load with Hugging Face Datasets

The default images configuration returns one row per unique image with a nested annotation list:

from datasets import load_dataset

dataset = load_dataset(
    "zrwang1211/SafeAtlas-VL",
    "images",
    streaming=True,
)
sample = next(iter(dataset["train"]))
image = sample["image"]
annotations = sample["annotations"]

The annotations configuration provides one row per safety annotation and does not include the image column:

annotations = load_dataset(
    "zrwang1211/SafeAtlas-VL",
    "annotations",
    streaming=True,
)
row = next(iter(annotations["train"]))

The two configurations can be joined with image_id.

Public schema

`images` configuration

Field	Type	Description
`image_id`	string	Split-local image identifier such as `train_img_0000000`.
`image`	Image	Encoded image data.
`annotations`	list	One or more safety annotations for the image.

Each nested annotation contains:

Field	Type	Description
`annotation_id`	string	Split-local annotation identifier.
`target`	string	One of `image`, `request`, or `response`.
`request`	nullable string	User request for request/response targets.
`response`	nullable string	Assistant response for response targets.
`safety_label`	string	Five-level ordinal safety label.
`category`	string	One of 15 harm categories, or `none` for `safe_core`.
`teacher_head`	nullable struct	Three teacher-head outputs; null for image targets.

`annotations` configuration

The flat view contains annotation_id, image_id, target, request, response, safety_label, category, and teacher_head.

The teacher-head struct contains:

qwen3guard: S, C, or U
guardreasoner_vl: 0 or 1
llamaguard4: 0 or 1

Labels

The ordered safety labels are:

safe_core
safe_leaning_disputed
boundary_uncertain
unsafe_leaning_disputed
unsafe_core

The 15 harm categories are dangerous information, defamation, erosion of trust in public information, false beliefs, fraud or deceptive action, illegal activities, influence operations, persuasion and manipulation, privacy, risky financial practices, security threats, toxic, trade and compliance, unfair, and violation of personal property.

Data organization

The release uses sharded Parquet files with the Hugging Face Image feature embedded in the schema. The images configuration contains image data and nested annotations, while the annotations configuration presents the annotations in tabular form.

Sensitive content warning

This dataset is designed for safety research and necessarily contains unsafe, offensive, sensitive, and potentially disturbing images, requests, and responses.

Intended use

Intended uses include multimodal safety moderation, ordinal risk assessment, red-teaming, evaluation, and safety alignment research. The dataset must not be used to facilitate harmful activity or to target individuals or protected groups.

Project links

Code: github.com/zrwang1211/SafeAtlas-VL
Models: SafeAtlas Guard collection

Citation

Citation metadata will be added after the arXiv identifier is assigned.

Downloads last month: 108

Total file size:

248 GB

Models trained or fine-tuned on zrwang1211/SafeAtlas-VL

Access SafeAtlas-VL