license: apache-2.0

Quixi AI Refusal Taxonomy

This is a comprehensive, production-grade refusal taxonomy based on the MLCommons Hazard Taxonomy and examples from Llama Guard, but significantly expanded and restructured for real-world deployment scenarios.

My goal is to train a classifier on the LiquidAI/LFM2-350M base model - to vastly outperform Llama Guard 4 12b.

Overview

This taxonomy provides a detailed classification system for identifying and categorizing harmful user prompts. It is empirically driven, with categories and examples reflecting actual threat patterns observed in production AI systems, rather than idealized academic structures. The framework is organized into 16 major categories and hundreds of granular subcategories to support the development of highly accurate and efficient safety models.

Dataset Statistics

• Total Categories: 16 major categories
• Subcategories: 300+ granular subcategories
• Example Prompts: 4,000+ real-world examples
• Coverage: Violent crimes, financial fraud, misinformation, jurisdiction-specific content, and more
• Languages: Primarily English with framework for multilingual expansion

Key Improvements Over Existing Taxonomies

• Empirically Driven: Categories are structured and weighted based on real-world threat volumes, not artificial balance
• Granular Categorization: Breaks down broad harms like "Violent Crimes" into specific, actionable subcategories (e.g., Mass Murder, Premeditated Murder, Crime of Passion)
• Comprehensive Examples: Each subcategory includes numerous real-world prompt examples, capturing diverse phrasing and user intent
• Production-Ready: Designed for direct use in operational training, evaluation, and policy enforcement pipelines
• Jurisdiction-Aware: First taxonomy to include region-specific content moderation requirements

*Target model size for classifiers trained on this taxonomy

New Categories Added

S15: Jurisdiction-Specific Content (China)

Addresses content moderation requirements for specific regions, including:

• Criticism of State Leadership and the Political System
• Discussions of Banned Historical Events
• Advocacy for Separatism or Independence
• Promotion of Banned Religious or Spiritual Groups
• Content on Human Rights Abuses and Political Dissent
• Evasion of State Censorship

S16: Misinformation and Harmful Conspiracies

Provides comprehensive coverage of modern misinformation patterns:

• Medical and Health Misinformation (e.g., Anti-vaccine, dangerous "cures")
• Harmful Conspiracy Theories
• Politically Motivated Hatred
• Historical Revisionism

Technical Goals

The primary goal of this taxonomy is to enable the training of highly efficient safety classifiers. The granular specificity is designed to allow smaller models to achieve superior performance through better training data quality rather than model scale.

Target Outcomes:

• Efficiency: ~500M parameter models achieving performance comparable to 7B+ models
• Accuracy: Superior coverage of real-world edge cases and emerging threats
• Adaptability: Framework supports rapid addition of new threat categories
• Deployment: Optimized for production environments with latency constraints

Structure

The taxonomy follows a clear hierarchical structure:

S[Number]: [Major Category]
├── [Subcategory Group]
│   ├── [Specific Subcategory]
│   │   ├── [Detailed prompt example]
│   │   └── [More examples...]
│   └── [Another Subcategory]
└── [Another Subcategory Group]

Quick Start

import json

# Load the taxonomy
with open('atpt_taxonomy.json', 'r') as f:
    taxonomy = json.load(f)

# Access specific categories
violent_crimes = taxonomy['S1: Violent Crimes']
misinformation = taxonomy['S16: Misinformation and Harmful Conspiracies']

# Example: Get all prompt examples for a subcategory
mass_murder_prompts = violent_crimes['Homicide and Murder']['Mass Murder']

Key Design Principles

1. Threat-Based: Categories reflect the frequency and severity of actual threats
2. Operational Focus: Built for production deployment, not just academic classification
3. Specificity: Granular enough to enable precise detection and minimize false positives/negatives
4. Extensibility: Structured to allow for easy addition of new categories as threats evolve
5. Cultural Awareness: Framework for jurisdiction-specific requirements
6. Asymmetric Reality: Reflects real-world threat distributions rather than artificial balance

Usage

This taxonomy is a foundational tool for:

• Training and fine-tuning content moderation classifiers
• Developing robust platform safety policies
• Creating high-quality evaluation and red-teaming datasets
• Designing annotation guidelines and training materials for human labelers
• Informing risk assessment frameworks for AI systems
• Building jurisdiction-specific content filters

Contributing

When adding new categories or examples:

• Base additions on observed, real-world patterns and data
• Maintain granular specificity and clear boundaries
• Include comprehensive prompt examples covering varied intent and phrasing
• Consider cross-category implications and ensure logical consistency
• Document empirical justification for new categories

Acknowledgments

This work builds upon the foundational research from:

• MLCommons for the original Hazard Taxonomy framework
• Meta AI for Llama Guard examples and methodology
• The broader AI safety research community for threat identification and categorization insights

License

This work is licensed under the Apache License, Version 2.0. You may obtain a copy of the License at:

http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

Citation

If you use this taxonomy in your research or product, please cite it as follows:

BibTeX Format:

@misc{atpt2024,
  author       = {{Your Name or Organization}},
  title        = {{Advanced Toxic Prompt Taxonomy (ATPT): A Production-Grade Framework for AI Safety}},
  year         = {2024},
  publisher    = {GitHub},
  journal      = {GitHub repository},
  howpublished = {\url{https://github.com/your-username/advanced-toxic-prompt-taxonomy}},
  note         = {Comprehensive taxonomy with 16 categories, 300+ subcategories, and 4,000+ examples}
}

Plain Text Format:

Your Name/Organization. (2024). Advanced Toxic Prompt Taxonomy (ATPT): A Production-Grade Framework for AI Safety. GitHub. Retrieved from https://github.com/your-username/advanced-toxic-prompt-taxonomy

For questions, issues, or collaboration opportunities, please open an issue or contact [your-contact-info].