Datasets:

danyarm
/

ckg-benchmark

license: cc-by-4.0
task_categories:
  - question-answering
  - text-retrieval
tags:
  - knowledge-graph
  - rag
  - retrieval
  - benchmark
  - llm
  - knowledge-representation
language:
  - en
pretty_name: CKG Benchmark
size_categories:
  - 10K<n<100K
configs:
  - config_name: domain_graphs
    data_files:
      - path: domains/*/learning-graph.csv
        split: train
  - config_name: results
    data_files:
      - path: results/*.csv
        split: train
  - config_name: queries
    data_files:
      - path: queries/*.jsonl
        split: train

CKG Benchmark

Pre-structured knowledge graphs outperform RAG by 4× F1 at 11× lower token cost — across 47 benchmarked domains.

System	Macro F1	Tokens/query	RDS	Run Cost
CKG	0.4709	269	0.00175	$7.81
RAG	0.1231	2,982	0.0000413	$76.23
GraphRAG	0.1200	3,450	0.0000452	$44.43

42× more intelligence per token than RAG. Zero hallucinations by construction.

Dataset Contents

domains/{domain}/learning-graph.csv   — structured DAG (ConceptID, ConceptLabel, Dependencies, TaxonomyID)
queries/queries_{domain}.jsonl        — 7,928 benchmark queries (T1–T5 types)
results/                              — per-system JSONL results + summary CSVs

Domain Library (52 total)

Benchmarked Educational Domains (47)

Domain	Category
algebra-1	Mathematics
asl-book	Language
automating-instructional-design	Education Technology
bioinformatics	Life Sciences
biology	Life Sciences
blockchain	Computer Science
calculus	Mathematics
chemistry	Natural Science
circuits	Engineering
claude-skills	AI / LLM
computer-science	Computer Science
conversational-ai	AI / LLM
data-science-course	Data Science
dementia	Healthcare
digital-citizenship	Social / Civic
digital-electronics	Engineering
ecology	Natural Science
economics-course	Social Science
ethics-course	Philosophy
fft-benchmarking	Signal Processing
functions	Mathematics
genetics	Life Sciences
geometry-course	Mathematics
glp1-obesity	Healthcare / Pharma
infographics	Design / Communication
intro-to-graph	Computer Science
intro-to-physics-course	Natural Science
it-management-graph	IT Management
learning-linux	Computer Science
linear-algebra	Mathematics
machine-learning-textbook	AI / Machine Learning
microsims	Education Technology
modeling-healthcare-data	Healthcare Analytics
moss	Biology / Botany
organizational-analytics	Business Analytics
personal-finance	Finance
pre-calc	Mathematics
prompt-class	AI / LLM
quantum-computing	Computer Science
reading-for-kindergarten	Education
signal-processing	Engineering
statistics-course	Data Science
systems-thinking	Systems Science
theory-of-knowledge	Philosophy
tracking-ai-course	AI / LLM
unicorns	Business / Finance
us-geography	Geography

Enterprise Domains (5, unbenchmarked — community contribution)

Domain	Category	Concepts
payer-formulary	Healthcare Payer Analytics	75
drug-interactions	Clinical Pharmacology	70
icd10-metabolic	Medical Coding	70
cpt-em-coding	Medical Billing	80
hipaa-compliance	Healthcare Compliance	75

Query Types

Type	Description	Example
T1	Entity lookup	"What is Composite Function?"
T2	Direct dependency	"What are the prerequisites for Implicit Differentiation?"
T3	Multi-hop path	"What is the prerequisite chain from Function to Taylor Series?"
T4	Category aggregate	"List all FOUND concepts"
T5	Cross-concept relationship	"How does Domain and Range relate to Inverse Function?"

Two-Track Design

Track 1 — McCreary Intelligent Textbook Corpus 44 open-source educational domains. Hand-authored learning-graph CSVs. STEM, Professional, Foundational.

Track 2 — Pipeline-Generated Commercial Domain GLP-1/Obesity pharmacology assembled from ClinicalTrials.gov API in one session. No expert curation. CKG F1 = 0.5298 — exceeds hand-curated average.

Key Finding: CKG improves with hop depth, RAG plateaus

hop depth	CKG F1	RAG F1
0	0.374	0.073
1	0.519	0.066
2	0.573	0.226
3	0.671	0.138
4	0.751	0.166
5	0.772	0.170

Novel Metrics

RDS (Retrieval Density Score) = F1 / tokens_consumed — intelligence per token
Hop-Depth F1 — multi-hop reasoning quality vs. chain length
CPCA — cost per correct answer

Citation

@misc{yarmoluk2026ckg,
  title={Benchmarking Knowledge Retrieval Architectures Across Educational
         and Commercial Domains: RAG, GraphRAG, and Compact Knowledge Graphs},
  author={Yarmoluk, Daniel and McCreary, Dan},
  year={2026},
  note={Pre-print in preparation. v0.6.2. Patent pending App #64/040,804.}
}

License

Dataset: CC BY 4.0
Source learning graphs: MIT (McCreary Intelligent Textbooks)
Enterprise domains: CC BY 4.0