license: cc-by-sa-4.0
tags:

• ai-safety
• alignment
• adversarial
• rlht
• red-teaming
• ethics
• reasoning
• llm
• owasp
• text-generation
  dataset_info:
  features:
  • name: ID
    dtype: string
  • name: Prompt
    dtype: string
  • name: Chosen Response
    dtype: string
  • name: Rejected Response
    dtype: string
  • name: RLHF Preference Pair
    dtype: string
    splits:
  • name: train
    num_bytes: # Calculate the size of GVD-1.json in bytes
    num_examples: 132 # Or the exact final count
    download_size: # Calculate the size of GVD-1.json in bytes
    dataset_size: # Calculate the size of GVD-1.json in bytes
    pretty_name: "Dormamu GVD-1: Guardrail Vulnerability Dataset for LLM Alignment"

Dormamu GVD-1: Alignment Benchmark Dataset

Introduction

Dormamu is an advanced AI alignment framework developed to enhance the safety and ethical robustness of large language models (LLMs). It extends the OWASP LLM Top 10 vulnerabilities with proprietary categories, such as DormamuX1 for existential risk simulations and DormamuX2 for cultural alignment testing.

The Guardrail Vulnerability Dataset version 1 (GVD-1) is a comprehensive collection of test cases designed to benchmark LLM alignment against prompt injections, sensitive data disclosures, supply chain vulnerabilities, and other critical risks.

Purpose

The GVD-1 dataset serves as an alignment benchmark to evaluate how well LLMs resist adversarial prompts, maintain ethical boundaries, and produce safe outputs. It supports Reinforcement Learning from Human Feedback (RLHF) training by providing structured preference pairs (chosen vs. rejected responses), enabling developers to fine-tune models for improved safety, fairness, and reliability in real-world applications.

Dataset Structure

The dataset is provided in JSONL format (in the GVD-1.jsonl file), where each line is a self-contained JSON object representing a single test case. Key fields include:

entry_id: Unique identifier (e.g., "DQ001").

prompt_source: Source of the prompt (e.g., "internal_team").

domain: Vulnerability category (e.g., "LLM01: Prompt Injection").

prompt: The adversarial or test prompt.

responses: An array containing two responses:

Chosen: The aligned, safe response.

Rejected: The misaligned, unsafe response.

human_feedback: Includes ranking, chosen response ID, and justification (e.g., multi-turn interactions).

For detailed explanations of all 132 categories, test descriptions, and Dormamu's unique innovations (e.g., narrative branching, bias neutralization), please refer to the metadata_guide.md file in this repository. The metadata.json file also provides a quick lookup for each test.

How to Load and Use the Dataset

You can easily load the dataset using the Hugging Face datasets library.

Example in Python:

from datasets import load_dataset

Load the dataset from the Hugging Face Hub

This automatically loads the GVD-1.jsonl file

dataset = load_dataset("Dormamu-Labs/Dormamu_GVD-1_Guardrail_Vulnerability_Dataset_for_LLM_Alignment")

Access a sample entry

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

This loads the data as a Hugging Face Dataset object, ready for tokenization, training, or analysis.

Citation

If you use this dataset in your research or applications, please cite it as follows:

@misc{dormamu_gvd1_2025,
title = {Dormamu GVD-1: Guardrail Vulnerability Dataset for LLM Alignment},
author = {Dormamu Team},
year = {2025},
url = {https://huggingface.co/datasets/Dormamu-Labs/Dormamu_GVD-1_Guardrail_Vulnerability_Dataset_for_LLM_Alignment}
}

License

This dataset is licensed under the Creative Commons Attribution-ShareAlike 4.0 International License (CC BY-SA 4.0).

For questions or contributions, please contact the Dormamu team at [email protected] or [email protected].