HALLMARK — Citation Hallucination Detection Benchmark

Anonymized for NeurIPS 2026 Datasets & Benchmarks Track double-blind review. Source code: https://anonymous.4open.science/r/hallmark/

HALLucination benchMARK evaluates citation verification tools on detecting hallucinated references in academic papers. The benchmark was motivated by the NeurIPS 2025 incident in which 53 accepted papers were found to contain fabricated citations that passed peer review.

TL;DR. 2,072 public bibliography entries (1,950 in dev_public+test_public, 122 in stress_test), labeled VALID or HALLUCINATED across 14 fine-grained hallucination types and 3 difficulty tiers, with 6 ground-truth sub-tests per entry.

At a glance

Split	Total	Valid	Hallucinated	Tier 1 / 2 / 3
dev_public	1,119	486	633	155 / 292 / 186
test_public	831	287	544	134 / 243 / 167
stress_test	122	1	121	— / 85 / 36
test_hidden	453	—	—	held out

• 14 hallucination types across 3 difficulty tiers (Easy / Medium / Hard) — see Taxonomy below.
• 6 sub-tests per entry: DOI resolution, title existence, author match, venue correctness, field completeness, cross-database agreement.
• Sources: DBLP-scraped real citations + perturbations + LLM-generated fakes + real-world hallucinations recovered from NeurIPS 2025 retracted papers (via GPTZero analysis).
• License: MIT.

Loading

from datasets import load_dataset

# Default config (all labels visible)
ds = load_dataset("hallmark-neurips2026/HALLMARK")

# Blind config (labels stripped — for leaderboard submissions)
blind = load_dataset("hallmark-neurips2026/HALLMARK", "blind")

The default config exposes the full schema with ground-truth labels. The blind config strips label, hallucination_type, difficulty_tier, explanation, subtests, and provenance fields — matching the format an evaluation tool would receive at inference time.

Schema

Each row is a citation entry with the following fields (default config):

Field	Type	Description
bibtex_key	str	Stable hex hash key (12 chars)
bibtex_type	str	inproceedings, article, misc, etc.
fields	dict[str, str]	BibTeX fields: author, title, year, doi, venue, pages, …
label	str	VALID or HALLUCINATED
hallucination_type	str / null	One of 14 types (null for VALID)
difficulty_tier	int / null	1 (easy), 2 (medium), 3 (hard)
explanation	str / null	Human-readable annotation rationale
generation_method	str	scraped, perturbation, llm_generated, adversarial, real_world, canary
source	str	Source pool identifier
source_conference	str / null	Originating venue if applicable
publication_date	str / null	ISO date for temporal contamination analysis
added_to_benchmark	str	ISO date the entry was added
subtests	dict[str, bool/null]	Six per-entry sub-test ground truths
raw_bibtex	str / null	Original BibTeX source if preserved
schema_version	str	Currently "1.0"

The blind config keeps only bibtex_key, bibtex_type, fields, and raw_bibtex.

Taxonomy

Tier	Type	Notes
1	fabricated_doi	DOI does not resolve
1	nonexistent_venue	Venue does not exist
1	placeholder_authors	"John Smith", "First Last", etc.
1	future_date	Publication year in the future
2	chimeric_title	Title combines fragments from multiple real papers
2	wrong_venue	Real paper, wrong venue
2	swapped_authors	Authors of a different real paper (a.k.a. author mismatch)
2	preprint_as_published	Cited as conference paper but only on arXiv
2	hybrid_fabrication	Real DOI but fabricated metadata
3	near_miss_title	Title close to a real paper, off by a few words
3	plausible_fabrication	LLM-generated entries that "look real" — sourced from real NeurIPS 2025 retractions
†	merged_citation	Two real citations fused into one
†	partial_author_list	Author list truncated past convention
†	arxiv_version_mismatch	Wrong arXiv version pin

† Theoretically-motivated stress-test types (present in all splits; concentrated in stress_test).

Repository layout

.
├── data/                          # Parquet (default config — full labels)
├── blind/                         # Parquet (blind config — labels stripped)
├── jsonl/                         # Original JSONL (lossless source of truth)
├── sources/                       # Provenance: pre-curation source pools
├── baseline_results/              # Pre-computed baseline predictions on dev_public
├── metadata.json                  # Per-split counts and distributions
├── source_mapping.json            # bibtex_key → source pool mapping
├── valid_entry_verification.json  # Verification trace for VALID entries
└── croissant_rai.json             # Responsible-AI metadata (Croissant RAI fields)

Evaluation

Use the accompanying code repository to run baselines and the official evaluator:

hallmark evaluate --split dev_public --baseline doi_only

hallmark evaluate --split dev_public \
    --predictions my_predictions.jsonl \
    --tool-name my-tool

Predictions follow this format:

{"bibtex_key": "abc123def456", "prediction": "HALLUCINATED", "confidence": 0.91}

Metrics reported

• Primary: Detection Rate, FPR, F1-Hallucination, Tier-weighted F1, Expected Calibration Error (ECE)
• Diagnostic: detect@k, source-stratified metrics, per-subtest accuracy, Plackett-Luce ranking (ONEBench-inspired)

---

Datasheet

This section follows the Datasheets for Datasets template (Gebru et al., 2021).

Motivation

• For what purpose was the dataset created? To enable rigorous, reproducible evaluation of citation-hallucination detection tools — a category of tools that received intense interest after the NeurIPS 2025 retraction incident exposed widespread fabricated citations passing peer review. No prior benchmark covered this task with controlled difficulty stratification and per-entry diagnostic sub-tests.
• Funded by. Anonymized for double-blind review.

Composition

• What do instances represent? Each instance is a single bibliography entry (BibTeX-style record) drawn from real papers, perturbed real entries, LLM-generated fakes, or real-world hallucinations.
• How many instances? 2,525 total (2,072 public + 453 hidden). See At a glance.
• What data does each instance consist of? Structured BibTeX fields (author, title, year, venue, doi, …), label, fine-grained hallucination type, difficulty tier, six sub-test ground truths, and provenance metadata.
• Is there a label or target? Yes: binary label ∈ {VALID, HALLUCINATED} plus hallucination_type (14 classes) and difficulty_tier.
• Are there missing values? subtests fields can be null when not applicable; raw_bibtex is null for synthetically constructed entries.
• Are relationships made explicit? Yes — source_mapping.json traces each bibtex_key to its source pool.
• Recommended train/dev/test split. dev_public for development, test_public for evaluation, stress_test for robustness analysis. test_hidden is reserved for a future leaderboard.

Collection process

• How was the data acquired? Real citations were scraped from DBLP. Perturbed entries were produced by deterministic transformations of real entries (e.g., year shift, author swap, venue substitution). LLM-generated entries were produced by prompting GPT, Claude, DeepSeek, Qwen, Gemini, and Mistral models to fabricate plausible citations. Real-world hallucinations were extracted from the GPTZero NeurIPS 2025 analysis (publicly released).
• Time frame. Real citations are dated 1994–2024 (publication_date field). Perturbations and LLM generations were produced in 2026.
• Were any ethical review or IRB processes used? No human subjects data; no IRB review required.

Preprocessing / cleaning / labeling

• Was preprocessing done? Yes: deduplication via bibtex_key hex hash, normalization of BibTeX field names, fuzzy-match guard against title leakage between splits.
• How were labels assigned?
  • scraped entries from DBLP — labeled VALID after positive verification (DOI resolves, title in DBLP/OpenAlex, year matches).
  • perturbation entries — labeled HALLUCINATED by construction with type recorded.
  • llm_generated and real_world entries — labeled HALLUCINATED after sub-test panel disagreement.
• Verification trace. valid_entry_verification.json records the resolver evidence used for each VALID entry.

Uses

• Intended uses.
  • Benchmarking citation-verification tools on hallucination detection.
  • Studying tool calibration (ECE) and tier-stratified failure modes.
  • Comparing LLM-as-judge approaches against rule-based and retrieval-based verifiers.
• Out-of-scope uses.
  • Not a training set for citation-generation models — using HALLMARK as supervised training data would contaminate downstream evaluation.
  • Not a measure of an LLM's general factuality outside the citation domain.
  • Not a substitute for end-to-end retraction review of any specific paper.

Distribution

• How is the dataset distributed? HuggingFace Datasets Hub, MIT license, with this card and a Croissant RAI metadata file.
• DOI / persistent identifier. Provided by HuggingFace upon publication.

Maintenance

• Who maintains the dataset? Anonymized for review. Post-acceptance, contact details will be added.
• Will the dataset be updated? Yes. Versioning follows semver-style tags on this repo (v1.0, v1.1, …). New hallucination instances are added quarterly; the hidden test split rotates annually to mitigate contamination. See the source repository's CHANGELOG.md.
• How can users contribute? A community contribution interface (hallmark contribute … CLI) is documented in the source repo. Submitted entries pass through the same six-sub-test verification pipeline before inclusion.
• Will old versions be retained? Yes — every tagged release is permanently retained as a HuggingFace revision.

---

Responsible AI considerations

A machine-readable Croissant RAI metadata file is shipped alongside this card as croissant_rai.json (validated against the MLCommons Croissant RAI spec).

Bias & representational concerns

• Source skew. Real citations are over-weighted toward NeurIPS, ICML, and ICLR — venues over-represented in DBLP scraping seeds. Tools tuned for medical or humanities citation conventions may underperform on this benchmark.
• Author-name distribution. Names follow real-world author distributions in CS publication corpora; tools relying on phonetic name heuristics may exhibit name-origin disparities.
• Language. English-only.

Harm-mitigation choices

• No PII beyond authorship metadata. All author names already appear in publicly indexed bibliographies.
• DOI prefix safety. Synthetic fabricated_doi entries use a fixed 10.99999/ prefix that is not registered with Crossref, preventing accidental hijack of any real DOI.
• No attack uplift. The benchmark is purely detection-oriented; we do not release prompts or pipelines that would meaningfully accelerate the production of plausible-looking fake citations beyond what is already in the public domain.

Limitations & known issues

• Contamination of LLM training data. LLM-based detectors trained after the NeurIPS 2025 retraction publicity may have seen overlapping content. We recommend reporting pre_cutoff vs post_cutoff slices via hallmark evaluate --temporal.
• Stress-test asymmetry. stress_test is overwhelmingly hallucinated (121/122). Use this split for recall stress-testing, not for FPR estimation.
• Tier boundary subjectivity. Tier assignment for near_miss_title is annotator-judged and may differ between reasonable annotators on borderline cases.

---

Reproducibility & code

• Source code: https://anonymous.4open.science/r/hallmark/ (fully executable; MIT license).
• Reproduction command: every split is regeneratable via python scripts/generate_new_instances.py --reproduce-v1 --seed 8042 against the source pools shipped under sources/.
• Pre-computed baseline results: baseline_results/ includes outputs for harc, bibtexupdater, and the LLM-agentic OpenAI baseline on dev_public, with checksums in manifest.json.

Provenance and contamination

• Temporal split: publication_date is recorded for every real-world entry, enabling pre/post-cutoff contamination analysis.
• Hidden split: test_hidden (453 entries) is not included in this release. It is held server-side for an upcoming leaderboard.
• Blind format: Use the blind config when reporting numbers from any tool that ingests the dataset directly — it removes any field a tool could shortcut on.

Citation

% Anonymized for NeurIPS 2026 D&B double-blind review.
% Citation BibTeX will be added to this card after acceptance / decision release.

License

MIT — see LICENSE in the source repository. Source citations themselves are bibliographic metadata in the public domain; perturbed and synthetic entries are released under MIT.