danyarm/ckg-benchmark

KRB — Knowledge Retrieval Benchmark v0.6.2

CKG outperforms RAG by 4× F1 at 11× lower token cost.

An open benchmark comparing Compressed Knowledge Graphs (CKG), RAG, and GraphRAG across 45 domains and 7,928 queries. Measures what production AI systems actually care about: accuracy, token cost, and dollar cost — all three.

📄 Full reference: graphifymd.com/benchmark
🏆 Live leaderboard: huggingface.co/spaces/danyarm/krb-leaderboard
📑 Paper: github.com/Yarmoluk/ckg-benchmark/blob/main/paper/main.pdf
🔖 DOI: 10.5281/zenodo.21222965

---

Results

Metric	CKG	RAG	GraphRAG
Macro-F1	0.471	0.123	0.120
Tokens / query	269	2,982	3,450
Cost / 1K queries	$7.81	$76.23	$44.43
5-hop reasoning F1	0.772	0.170	—
Hallucination	0 by construction	uncontrolled	uncontrolled

Reasoning depth

CKG F1 improves monotonically with hop depth. RAG plateaus and declines.

Hop depth	CKG F1	RAG F1
0	0.374	0.148
1	0.425	0.131
2	0.480	0.119
3	0.551	0.108
4	0.648	0.096
5	0.772	0.170

At 5-hop reasoning depth: CKG 0.772 vs RAG 0.170 — 4.5× gap.

---

What's in the dataset

• 14,043 rows across 45 benchmarked domains
• 7,928 benchmark queries spanning 5 types:
  • T1: Entity lookup (single-hop)
  • T2: Dependency traversal
  • T3: Multi-hop path queries
  • T4: Aggregate queries
  • T5: Relationship queries
• Per-row fields: query, answer, domain, hop_depth (0–5), query_type (T1–T5)
• Results JSONL with per-system F1, token count, and dollar cost

Domains covered

Mathematics · Physics · Chemistry · Biology · Computer Science · Economics · Philosophy · Psychology · Sociology · History · Law · Medicine · AI/ML · Data Science · Software Engineering · Financial Regulation · Life Sciences · Healthcare · Engineering · and 27 more.

---

Use the dataset

from datasets import load_dataset

ds = load_dataset("danyarm/ckg-benchmark")

for row in ds["train"]:
    query        = row["query"]
    hop_depth    = row["hop_depth"]    # 0–5
    ground_truth = row["answer"]
    domain       = row["domain"]

    # Run your retrieval system here
    # Track: F1 score, token count, dollar cost

# Submit results to the leaderboard:
# huggingface.co/spaces/danyarm/krb-leaderboard

Evaluate your system

from datasets import load_dataset
from sklearn.metrics import f1_score

ds = load_dataset("danyarm/ckg-benchmark", split="train")

predictions = []
references  = []

for row in ds:
    pred = your_retrieval_system(row["query"])   # your system here
    predictions.append(pred)
    references.append(row["answer"])

# Compute token-normalized F1
macro_f1 = compute_f1(predictions, references)
print(f"Macro-F1: {macro_f1:.3f}")
print(f"Baseline — RAG: 0.123 | GraphRAG: 0.120 | CKG: 0.471")

---

Novel metrics introduced

Metric	Description
RDS (Retrieval Density Score)	Intelligence per token consumed — combines F1 and token cost
Hop-Depth F1	F1 stratified by reasoning chain length (0–5 hops)
CPCA	Cost-Per-Correct-Answer — dollar cost normalized by accuracy

---

Cite

@dataset{yarmoluk2026krb,
  title    = {KRB: Knowledge Retrieval Benchmark v0.6.2},
  author   = {Yarmoluk, Daniel},
  year     = {2026},
  doi      = {10.5281/zenodo.21222965},
  url      = {https://huggingface.co/datasets/danyarm/ckg-benchmark},
  license  = {CC BY 4.0},
  note     = {CKG Macro-F1 0.471 vs RAG 0.123 across 45 domains,
              7,928 queries. 11× token reduction, ~10× cost reduction.}
}

---

About

Built by Daniel Yarmoluk at Graphify.md.
Compressed Knowledge Graphs (CKGs) are a patent-pending methodology for encoding domain knowledge as typed, traversable graphs — served natively via MCP.

License: CC BY 4.0 · Domains: 45 benchmarked · Queries: 7,928 · Version: v0.6.2