prithivMLmods/ImageShield-Guardrail-Traces

This dataset contains sensitive visual content intended for training Image Guard, Image Filter, and Image Shield models. Due to the sensitive nature of the data, access is gated and requires manual review. Access may be approved for users who provide a valid research, safety, or development purpose.

ImageShield-Guardrail-Traces

A multimodal image safety dataset containing visual inputs and structured guardrail classification traces. The dataset is designed for image understanding, safety classification, and explanation-oriented multimodal model training. ImageShield-Guardrail-Traces contains 5,000 image-text samples. Each sample pairs an image with a corresponding text output that describes the visual content, explains the relevant safety classification, and provides structured Boolean labels across six guardrail categories.

The dataset follows a consistent three-part output format:

• DESCRIPTION: A concise, objective description of the image.
• EXPLANATION: An explanation of the safety classification based on the visible content.
• GUARDRAIL: A JSON object containing Boolean labels for categories C1 through C6.

Dataset Details

Property	Value
Dataset	ImageShield-Guardrail-Traces
Author	prithivMLmods
Samples	5,000
Task	Image-to-Text
Modalities	Image, Text
Language	English
Features	Image, Output
Format	Image-text pairs
License	Apache-2.0

Dataset Structure

The dataset contains two columns.

Column	Type	Description
Image	Image	Input image to be analyzed
Output	String	Structured description, explanation, and guardrail classification

Each output follows a consistent structure to support supervised fine-tuning and multimodal safety classification workflows.

Guardrail Categories

The output uses six Boolean classification labels.

Label	Category	Description
C1	Sexual Content	Sexual content, nudity, or explicit sexual activity
C2	Harassment & Bullying	Targeted bullying, degradation, humiliation, or harassment
C3	Threats, Violence & Harm	Physical violence, threats, or depictions of harm
C4	False & Deceptive Information	False or misleading information and deceptive content
C5	Illegal/Regulated Activities	Content involving illegal or regulated activities
C6	Hateful Content & Extremism	Hate-related content, extremist material, or associated symbols

A category is marked true when the content meets the applicable classification criteria. Otherwise, it is marked false.

These labels are not mutually exclusive. Multiple categories may be marked true when the content clearly meets more than one category's criteria.

Sample Trace

Input: An image depicting two individuals physically grappling in what appears to be a school hallway.

Output:

DESCRIPTION: Two individuals are physically grappling or fighting in what appears to be a school hallway.

EXPLANATION: The image displays a physical confrontation where two individuals are grappling in a public indoor space. This content is classified under Violence & Harm as it features direct physical struggle and physical force.

GUARDRAIL: {"C1": false, "C2": false, "C3": true, "C4": false, "C5": false, "C6": false}

The example illustrates how an image is converted into a concise visual description, a category-based explanation, and a machine-readable guardrail classification.

Intended Use

The dataset may be useful for:

• Multimodal safety classification: Training models to identify safety-relevant visual content.
• Supervised fine-tuning (SFT): Teaching vision-language models to produce structured classification outputs.
• Visual content moderation: Exploring image-based safety detection and content filtering.
• Explanation generation: Training models to provide concise, evidence-based explanations for their classifications.
• Structured output generation: Improving consistency in generating predefined fields and Boolean JSON labels.
• Guardrail evaluation: Evaluating model predictions against annotated category labels.

Loading the Dataset

Load the dataset directly from Hugging Face using the datasets library.

from datasets import load_dataset

dataset = load_dataset(
    "prithivMLmods/ImageShield-Guardrail-Traces"
)

print(dataset)
print(dataset["train"][0])

Access an individual image and its corresponding output:

sample = dataset["train"][0]

image = sample["Image"]
output = sample["Output"]

print(output)

The Image column contains the visual input, while Output contains the associated classification trace.

Training Considerations

For supervised fine-tuning, the Image column can be used as the visual input and Output as the target text.

Depending on the training objective, the output can be used as a complete target sequence or parsed into separate components:

1. DESCRIPTION for visual description.
2. EXPLANATION for classification justification.
3. GUARDRAIL for structured category prediction.

For classification-only training, the Boolean labels can be extracted from the JSON object. For explanation-oriented training, the complete output can be retained.

The output format should be validated before training to ensure that all six category keys are present, their values are Boolean, and the JSON is valid.

Limitations

• The dataset contains model-generated classification traces, which may include incorrect interpretations or inconsistent category assignments.
• Visual context may be insufficient to determine intent, the circumstances surrounding an event, or whether an incident constitutes targeted harassment.
• Some categories may overlap, and classification quality depends on the definitions and annotation criteria used.
• The presence of a depicted activity does not necessarily establish that the activity is illegal, harmful in context, or directed at a particular individual.
• The dataset should not be treated as a definitive source of safety judgments without independent validation.

For research and deployment, evaluate the dataset against a manually reviewed validation set and measure category-wise precision, recall, and false-positive rates.

License

This dataset is published under the Apache License 2.0.

Users should independently verify the licensing and usage rights of the source images before redistribution or commercial use.

Citation

If you use this dataset in your research or projects, please reference the dataset repository:

@misc{imageshield_guardrail_traces,
  author       = {prithivMLmods},
  title        = {ImageShield-Guardrail-Traces},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/prithivMLmods/ImageShield-Guardrail-Traces}}
}

Acknowledgements

Developed by prithivMLmods as part of the ImageShield collection of multimodal safety and guardrail datasets.