AILuminate Taxonomy Mapping Rationale

Explanation of how 117 native harm class labels map to the MLCommons AILuminate v1.0 taxonomy

taxonomy Last updated: February 6, 2026

AILuminate Taxonomy Mapping Rationale

This document explains the rationale behind the mapping of 117 native harm class labels into the 12 MLCommons AILuminate v1.0 hazard categories.

1. Philosophy: Content vs. Method

Our project uses a Dual Taxonomy model:

  1. Content Dimension (AILuminate): What is the harmful outcome? (e.g., Violent Crime).
  2. Method Dimension (FERT labels.intent.*): How was the failure induced? (e.g., Persona Hijack).

This mapping addresses the Content Dimension to ensure our datasets are interoperable with industry standards.

2. Category Rationale & Edge Cases

Below is a short rationale for each of the 12 AILuminate categories, followed by key edge cases.

2.0 Category-by-Category Rationale (All 12)

  • Violent Crimes: Direct physical harm to people (assault, murder, gore).
  • Indiscriminate Weapons (CBRN): Chemical/biological/radiological/nuclear or explosive weaponization.
  • Suicide and Self-Harm: Self-directed harm, suicide methods, eating-disorder exploitation.
  • Nonviolent Crimes: Fraud, theft, cybercrime, illegal access, nonviolent wrongdoing.
  • Sex-Related Crimes: Sexual assault, coercion, trafficking (non-consensual).
  • Child Sexual Exploitation: CSAM, grooming, child abuse content.
  • Hate: Hate speech, discrimination, harassment based on protected traits.
  • Defamation: Libel, misinformation/disinformation about individuals/groups.
  • Privacy: PII leakage, doxxing, surveillance, privacy violations.
  • Intellectual Property: Piracy, copyright infringement, IP theft.
  • Sexual Content: Consensual adult sexual content (non-criminal).
  • Specialized Advice: Medical/legal/financial/political advice requiring licensure.

2.1 Non-sexual Explicit Content → Violent Crimes

  • Mapping: Explicit content (non-sexual) → Violent Crimes
  • Rationale: Many public datasets group “gore,” “mutilation,” and “extreme violence” under “explicit content.” However, AILuminate reserves the Sexual Content category strictly for sexualized material. We map these to Violent Crimes as they represent physical hazards rather than contextual sexual hazards.

2.2 Animal Abuse → Nonviolent Crimes

  • Mapping: Animal abuse → Nonviolent Crimes
  • Rationale: AILuminate lacks a specific category for animal cruelty. While animal abuse often involves physical violence, we categorize it under Nonviolent Crimes (as a property or ethical violation) to keep the Violent Crimes category focused on human physical hazards, per AILuminate’s typical usage.

2.3 Military Use → Indiscriminate Weapons (CBRN)

  • Mapping: Military instructions → Indiscriminate Weapons (CBRN)
  • Rationale: Prompts requesting tactical military instructions or battlefield logistics are mapped to the weapons category, as they represent the application of indiscriminate force in a non-permissive context.

2.4 Unmapped Datasets (NULL Mapping)

The following datasets are mapped to NULL for AILuminate categories:

  • DAN-In-The-Wild: Focuses on Attack Methods (DAN, STAN, etc.) rather than specific harms.
  • ToxicChat (jailbreak subset): Focuses on the act of jailbreaking itself.
  • TDC2023-RedTeaming: Competition prompts that often test generic helpfulness or boundary-testing rather than specific hazard categories.
  • AdvBench: Mixed harmful prompts without a stable per-prompt harm label.
  • WildJailbreak: Aggregated jailbreak prompts; technique-focused rather than harm-focused.
  • StrongREJECT: Mixed refusals without content category labels.

3. AILuminate Domain Breakdown

AILuminate DomainFocus Area in FERT
Physical HazardsViolent crimes, weapons, self-harm. Critical for embodied robotics.
Nonphysical HazardsHate speech, defamation, non-violent crimes.
Contextual HazardsPrivacy violations, IP theft, specialized medical/legal advice.