data/benchmarks/README.md

metadata

license: cdla-permissive-2.0
task_categories:
  - text-classification
  - feature-extraction
  - question-answering
language:
  - en
tags:
  - legal
  - eu-ai-act
  - regulation-2024-1689
  - knowledge-graph
  - deontic-logic
  - nist-ai-rmf
  - iso-42001
  - gdpr
  - neuro-symbolic
size_categories:
  - n<1K
dataset_info:
  features:
    - name: id
      dtype: string
    - name: article_number
      dtype: int64
    - name: paragraph_number
      dtype: string
    - name: subject
      dtype: string
    - name: modality
      dtype: string
    - name: predicate
      dtype: string
    - name: object
      dtype: string
    - name: source_text
      dtype: string
    - name: sha256_hash
      dtype: string
    - name: cross_references
      sequence: string
    - name: enforcement_tier
      dtype: string
    - name: shacl_shape_ref
      dtype: string
  splits:
    - name: train
      num_bytes: 30248
      num_examples: 36

🏛️ EU AI Act Normative Deontic Triples & Knowledge Graph

Formal Symbolic Regulatory Knowledge Base & Multi-Framework Crosswalk

Regulation (EU) 2024/1689 (Artificial Intelligence Act)

---

📌 Executive Summary

The EU AI Act Normative Deontic Triples dataset provides a rigorous, machine-verifiable, symbolic representation of Regulation (EU) 2024/1689. Built using the knowledge engineering methodology established in gitmodelmujtaba/gdpr-normative-triples, this benchmark translates dense legal prose into formal Subject-Modality-Predicate-Object tuples grounded in Deontic Logic (OBLIGATION, PROHIBITION, PERMISSION, EXEMPTION).

Each triple is bound to an exact EUR-Lex statutory quote, anchored with a cryptographically verifiable SHA-256 hash, cross-mapped bidirectionally to NIST AI RMF 1.0, ISO/IEC 42001:2023, and GDPR (EU 2016/679), and coupled with an executable Cytoscape/NetworkX knowledge graph topology.

---

🏛️ 5-Tier Dataset Architecture

data/benchmarks/
├── eu_ai_act_normative_triples.json     # Tier 1: Canonical Deontic Triples with SHA-256 digests
├── eu_ai_act_knowledge_graph.json       # Tier 2: Cytoscape & NetworkX graph topology (elements.nodes/edges)
├── domain_data_dictionary.json          # Tier 3: Controlled legal taxonomy, actor roles & risk tiers
├── rules/
│   ├── ai_act_fine_guidelines.json      # Tier 4a: Article 99 administrative fine tiers (35M€/7%, 15M€/3%)
│   └── cross_regulatory_frameworks.json # Tier 4b: Multi-framework ontology mappings (NIST, ISO, GDPR)
├── provenance_ledger.json               # Tier 5: Cryptographic Merkle provenance ledger
└── conformity_ground_truth_benchmark.jsonl # Ground-truth evaluation cases

---

⚖️ Deontic Logic Modal Specification

Every regulatory statement is classified under formal deontic logic:

Deontic Modality	Formal Meaning	Statutory Markers	Example Clause	Fine Exposure
PROHIBITION	Forbidden practice; non-compliance is strictly unlawful	"shall not", "prohibited", "unlawful"	Article 5(1)(c) Social Scoring	Up to 35M€ or 7% global turnover
OBLIGATION	Mandatory positive duty	"shall", "must", "is required to"	Article 9 Continuous Risk Management	Up to 15M€ or 3% global turnover
PERMISSION	Discretionary statutory right	"may", "is entitled to"	Article 10(5) Sensitive data for bias correction	N/A
EXEMPTION	Statutory safe harbor or carve-out	"shall not apply to", "derogation"	Article 2(3) Exclusively military / defense AI	Safe Harbor

---

📊 Dataset Statistics & Coverage

• Total Deontic Triples: 36
• Total Knowledge Graph Nodes: 105
• Total Relational Edges: 222
• Articles Grounded: Article 5 (Prohibitions), Article 9 (Risk Management), Article 10 (Data Governance & Bias Mitigation), Article 11 (Annex IV Technical Docs), Article 12 (Automatic Logging), Article 13 (Transparency), Article 14 (Human Oversight & Kill-Switch), Article 15 (Accuracy, Robustness & Cybersecurity), Article 26 (Deployer Duties), Article 27 (FRIA), Article 50 (Generative AI & Deepfakes), Article 51 (GPAI Systemic Risk > 10^25 FLOPs), Article 53 & 55 (GPAI Red-Teaming), Article 99 (Penalties).
• Cross-Framework Mappings: 22 bidirectional links to NIST AI RMF 1.0 (GOVERN, MAP, MEASURE, MANAGE), ISO/IEC 42001:2023, and GDPR Articles 22, 25, 32, 35.

---

💻 Quickstart: Loading in Python

1. Load via Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("gitmodelmujtaba/eu-ai-act-normative-triples", split="train")
print(dataset[0])
# {
#   "id": "EU_AIA_TRIPLE_003",
#   "article_number": 5,
#   "paragraph_number": "1(c)",
#   "modality": "PROHIBITION",
#   "predicate": "shallNotPlaceOnMarketOrPutIntoService",
#   "object": "Social_Scoring_AI_System",
#   "enforcement_tier": "TIER_1_PROHIBITED_AI",
#   ...
# }

2. Load Knowledge Graph into NetworkX

import json
import networkx as nx

with open("data/benchmarks/eu_ai_act_knowledge_graph.json", "r", encoding="utf-8") as f:
    kg = json.load(f)

G = nx.DiGraph()
for node in kg["elements"]["nodes"]:
    G.add_node(node["data"]["id"], **node["data"])

for edge in kg["elements"]["edges"]:
    G.add_edge(edge["data"]["source"], edge["data"]["target"], **edge["data"])

print(f"Graph loaded: {G.number_of_nodes()} nodes, {G.number_of_edges()} edges")

3. Load Cytoscape.js Format in Web UI

The eu_ai_act_knowledge_graph.json contains a direct elements dictionary compatible with Cytoscape.js:

const cy = cytoscape({
  container: document.getElementById('cy'),
  elements: data.elements,
  style: [
    { selector: 'node[category="prohibited_practice"]', style: { 'background-color': '#ef4444', 'label': 'data(label)' } },
    { selector: 'node[category="regulatory_requirement"]', style: { 'background-color': '#3b82f6', 'label': 'data(label)' } },
    { selector: 'edge[modality="PROHIBITION"]', style: { 'line-color': '#ef4444', 'target-arrow-color': '#ef4444', 'target-arrow-shape': 'triangle' } }
  ]
});

---

🔒 Cryptographic Provenance Ledger

Every file in this benchmark is hashed with SHA-256 and committed to provenance_ledger.json.

• Merkle Root Digest: cbf58e52a6f9066ec12826e4fbb7aa267c5a32180566fc0afe47e27eaefe1d08
• Genesis Statutory Text: EUR-Lex CELEX:32024R1689 (Official Journal of the European Union, L 2024/1689)
• Hash Integrity Guarantee: sha256(source_text) allows zero-hallucination downstream citation auditing.

---

📑 Citation & BibTeX

@dataset{eu_ai_act_normative_triples_2026,
  author       = {Mujtaba Hussain},
  title        = {EU AI Act Normative Deontic Triples & Knowledge Graph},
  year         = {2026},
  publisher    = {Hugging Face},
  version      = {2.0.0},
  url          = {https://huggingface.co/datasets/gitmodelmujtaba/eu-ai-act-normative-triples}
}