cloverxion/image-moderation-dataset

Image Moderation Dataset (Binary Classification: Safe vs. Unsafe)

CRITICAL WARNING: Contains Highly Sensitive, Explicit, and Uncensored Content Please be advised that while this dataset includes a general 'Safe' class, the 'Unsafe' class contains raw, completely uncensored human nudity, explicit adult content, graphic violence (gore), and weaponry. This repository is strictly intended for institutional academic research, AI safety engineering, and automated content moderation benchmarking. Viewer discretion and rigorous data handling protocols are absolutely mandatory.

---

1. Dataset Taxonomy and Architecture

This dataset serves as the foundational binary classification baseline (Safe vs. Unsafe) for image moderation systems. It is packaged within a ZIP archive to maintain strict file structure integrity and metadata organization, comprising a balanced distribution of over 15,000 images per class (30,000+ images in total).

Class Structure:

1. SAFE (15,000+ Images): General imagery deemed Safe for Work (SFW), including standard human interactions, everyday objects, and safe environments.
2. UNSAFE (15,000+ Images): Hazardous and restricted imagery. This class aggregates severe and uncensored content, including explicit human nudity, severe gore, graphic violence, and illegal weaponry.

Crucial Technical Note: Researchers are strongly advised to verify the class order and index mapping within their respective data loaders prior to initiating model training to prevent critical classification misalignment (silent errors).

Exclusion of Illegal Material: Despite the raw and explicit nature of the 'Unsafe' class, this dataset strictly adheres to international legal boundaries regarding cyber-exploitation. It does not contain Child Sexual Abuse Material (CSAM) or any content that violates fundamental human rights laws.

2. Comprehensive Terms of Use and Legal Disclaimer

By accessing, downloading, cloning, or utilizing this dataset in any capacity, you legally acknowledge and unconditionally agree to the following strict provisions:

A. Data Provenance and Copyright Exemption (Internet Disclaimer)

All materials within this dataset were programmatically aggregated from open, public internet sources solely to establish a real-world safety benchmark. The dataset creator does not claim intellectual property rights or copyright ownership over any individual images, did not produce the content, does not monetize this repository, and provides the data strictly "as-is" utilizing the principles of 'Fair Use' for non-profit academic research.

B. Absolute Prohibition of Commercial Use

This dataset is shielded by a strict non-commercial license. You are expressly prohibited from utilizing this data, or any AI model weights, embeddings, or APIs trained directly or indirectly on this data, for commercial SaaS endpoint monetization, paid products, or any profit-generating platforms. Any commercial application requires distinct, explicit written authorization from the respective original copyright holders.

C. Permissive Scope for Generative AI and Academic Research

The dataset may be utilized to train, fine-tune, or evaluate generative AI frameworks (e.g., Diffusion Architectures, GANs, Vision-Language Models) and visual reconstruction systems, provided the scope is strictly confined to non-profit academic research. Permitted use cases include:

• Safety Alignment and Machine Unlearning: Utilizing the stark contrast between the 'Safe' and 'Unsafe' classes to research Concept Erasure methodologies, forcing AI models to structurally reject the generation of unsafe concepts.
• Adversarial Red-Teaming: Stress-testing multimodal systems and moderation filters against adversarial edge cases and visual jailbreaks.
• Anatomical and Structural Studies: Research focusing on the accurate procedural generation of human anatomy or structural damage for non-profit educational or medical science purposes.
• Algorithmic Defense Development: Engineering robust automated moderation heuristics and content provenance tracking systems.

D. Strict Prohibition of Malicious Weaponization

While generative experimentation is conditionally permitted under Section C, the resulting AI weights or generated visual outputs must never be weaponized. Users are strictly prohibited from utilizing insights derived from this dataset for:

• Generating and distributing non-consensual synthetic media (e.g., targeted deepfakes or face-swapping).
• Creating synthetic explicit pornography for public distribution, blackmail, or harassment.
• Producing photorealistic misinformation intended for financial fraud or legal deception.

E. Data Containment, Handling, and No Redistribution

Given the hazardous nature of the 'Unsafe' class, researchers must adhere to strict data security protocols:

• No Mirroring: You are strictly prohibited from re-uploading, mirroring, distributing, or sharing this dataset archive (partially or entirely) to alternative public or private platforms, torrents, or cloud buckets. All distribution must remain centralized here.
• Secure Storage: The dataset must be stored locally or on secure, access-controlled institutional servers.

F. Total Legal Indemnity and Assumption of Risk

Because this data is sourced from the public internet purely to advance the field of AI safety, the dataset creator, contributors, and hosting platform are fully indemnified from any legal actions, copyright infringement claims, privacy violations, psychological impacts, or damages resulting from the handling or misuse of technology developed by third parties utilizing these data samples. All risks associated with the possession, processing, and application of this dataset are borne entirely by the user.

3. Related Infrastructure

If your research requires deeper granular classification within the hazardous spectrum (e.g., isolating Gore from NSFW instead of grouping them into a single 'Unsafe' class), please refer to our complementary multi-class specialist dataset: cloverxion/image-moderation-specialist

4. Citation

If you utilize this dataset in your academic research, safety alignment papers, or benchmarking evaluations, please implicitly cite this repository using the following BibTeX format:

@misc{cloverxion2026imagemoddataset,
  author       = {XoneMi},
  title        = {Image Moderation Dataset (Binary Classification: Safe vs. Unsafe)},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/cloverxion/image-moderation-dataset}},
  note         = {Accessed: 2026}
}