IntegriRef-Bench / README.md
Geoffrey-Wang's picture
Upload folder using huggingface_hub
4fea62c verified
metadata
license: mit
task_categories:
  - text-classification
  - fact-checking
language:
  - en
tags:
  - citation-verification
  - hallucination-detection
  - scientific-integrity
  - bayesian-scoring
  - reference-checking
size_categories:
  - 10K<n<100K

IntegriRef-Bench

Benchmark dataset for multi-level reference integrity verification, accompanying the IntegriRef framework.

Dataset Splits

Split Rows Description
reference_verification 1,926 Golden benchmark + crawled verification cases (hallucinated, real, chimera, retracted)
signal_unit_tests 403 Per-signal unit tests for 14 Bayesian signal types
l1_intent 20 Citation intent classification test pairs
l2_nli 20 Claim-evidence NLI alignment test pairs
graph_anomaly 3,030 Citation graph anomaly cases (rings, temporal, orphan clusters)
retracted_papers 6,391 Retracted papers from Crossref + PubMed with real controls
Total 11,790

Usage

from datasets import load_dataset

ds = load_dataset("Geoffrey-Wang/IntegriRef-Bench")

# Or load a specific split
ref_ver = load_dataset("Geoffrey-Wang/IntegriRef-Bench", data_files="reference_verification.jsonl")

Sources

  • Retracted papers: Crossref update-to retraction markers + PubMed retraction notices
  • Hallucinated references: Programmatically generated with verified non-existence
  • Chimera references: Real DOIs paired with swapped metadata
  • Graph anomaly cases: Documented citation cartels (Brazilian 2009-2013, Ji-Huan He, IOP 2024)
  • Temporal anomalies: OpenAlex citation graph analysis
  • Statistical cases: GRIM test + statcheck from PMC Open Access full texts

Citation

@inproceedings{integriref2026,
  title={IntegriRef: A Five-Layer Bayesian Framework for Cross-Domain Reference Integrity Verification},
  author={Anonymous},
  booktitle={KnowFM Workshop at ACL},
  year={2026}
}

License

MIT