citation-integrity-bench
收藏资源简介:
Citation Integrity Bench 是一个专门用于评估大语言模型生成虚假引用(幻觉引用)检测器的标注数据集。其创建背景源于虚假引用对学术文献的污染问题,旨在为检测工具提供标准化的评估基准。数据集包含两个核心部分:1) 一个包含 256 个标注案例的引用完整性评估集,每个案例对应一个参考文献条目,并标注为四种类型之一:real(真实且元数据正确)、real_partial(真实但元数据不完整)、poisoned_metadata(真实作品但作者或年份等关键元数据错误,模拟常见LLM错误)、fabricated(完全虚构的作品)。其中,虚构和元数据被“下毒”的条目是人工合成的,仅用于评估目的,不针对任何真实论文。数据以 JSONL 和 BibTeX 格式提供,包含 id, label, title, authors, year, doi 等多个字段。2) 一个包含 285 个主张-引用对的子集,用于评估引用在词汇层面是否支持给定的主张。数据集的主要用途是评估检测器在判断引用是否存在及其元数据完整性方面的性能,使用者可以通过比较检测器输出与数据集中预定义的 expected_status_family(如 verified, metadata_mismatch, not_found 等)来评分。数据集采用 CC0-1.0 许可证发布。
Citation Integrity Bench is a labeled dataset specifically designed for evaluating large language model-generated false citation (hallucination citation) detectors. Its creation background stems from the issue of false citations polluting academic literature, aiming to provide a standardized evaluation benchmark for detection tools. The dataset consists of two core parts: 1) A citation integrity evaluation set containing 256 labeled cases, each corresponding to a reference entry and annotated as one of four types: real (real with correct metadata), real_partial (real but with incomplete metadata), poisoned_metadata (real work but with key metadata errors such as author or year, simulating common LLM errors), and fabricated (completely fabricated work). Among these, the fabricated and metadata-poisoned entries are artificially synthesized, solely for evaluation purposes and not targeting any real papers. The data is provided in JSONL and BibTeX formats, including fields such as id, label, title, authors, year, and doi. 2) A subset containing 285 claim-citation pairs, used to evaluate whether citations lexically support given claims. The main purpose of the dataset is to assess the performance of detectors in judging the existence of citations and the integrity of their metadata. Users can score by comparing detector outputs with predefined expected_status_family (e.g., verified, metadata_mismatch, not_found) in the dataset. The dataset is released under the CC0-1.0 license.
数据集概述:Citation Integrity Bench
- 数据集名称:Citation Integrity Bench
- 许可证:CC0-1.0(公共领域奉献)
- 语言:英语、中文、德语
- 任务类别:文本分类、其他
- 标签:引用、文献计量学、幻觉检测、学术、评估
- 数据集规模:少于1000条(n<1K)
数据集内容
该数据集是一个精心标注的评估集,包含 256 个标注案例,用于检测引文存在性和元数据完整性。旨在帮助检测工具识别大语言模型(LLM)虚构的参考文献。
标签含义
| 标签 | 含义 |
|---|---|
real |
真实作品,核心元数据正确 |
real_partial |
真实作品,元数据不完整但无矛盾 |
poisoned_metadata |
真实标题/作品,但作者或年份错误(常见的LLM失败模式) |
fabricated |
不存在该作品(应为 not_found) |
文件结构
data/citations.jsonl:每条引用一个JSON对象data/citations.bib:相同数据集的BibTeX格式,适用于CLI工具
字段说明
id, label, title, authors, year, doi, arxiv_id, venue, notes, expected_status_family
预期用途
通过比较检测结果与 expected_status_family 来评分检测工具。expected_status_family 取值及其映射关系:
verified_or_mismatch_ok→verified或轻微metadata_mismatchverified→verifiedmetadata_mismatch_or_ambiguous→metadata_mismatch或ambiguousmetadata_mismatch_or_not_found→metadata_mismatch,ambiguous, 或not_foundnot_found→not_foundretracted→retracted
额外子集:Claim Support Mini-Bench
- 文件:
data/claims.jsonl - 规模:285 个 声明-引用对
- 用途:用于词汇层面声明-支持评分(
supported_or_weak/unsupported_or_weak家族) - 工具:配套于
citesure claims-file命令使用
基线性能
使用 citesure 工具在数据集上的基线结果:250/250(基于美国判例法,通过CourtListener;包含CJK标题)。





