fever-deception-source
收藏资源简介:
FEVER Deception Source Dataset 是一个基于 FEVER 数据集构建的诚实与欺骗性事实核查响应对数据集,旨在为欺骗检测探针提供训练数据。每个样本包含来自 FEVER 的声明(带有真实标签 SUPPORTS 或 REFUTES)、一个诚实完成(正确得出结论)和一个欺骗性完成(故意使用看似合理的推理得出错误结论)。数据集通过生成模型 `mlabonne/gemma-3-27b-it-abliterated` 和评判模型 `Qwen/Qwen3-30B-A3B` 生成,并基于 FEVER v1.0 训练集中的 SUPPORTS 和 REFUTES 声明(排除 NEI)。诚实完成基于维基百科文章的证据句子。数据集包含两个批次(batch1_seed42 和 batch2_seed43),共 38,927 行数据,来自 34,375 个独特声明(批次间声明重复率为 23.4%)。统计数据包括声明类型分布(SUPPORTS 和 REFUTES 各约 50%)、欺骗风格(explicit_lie 占 65.7%,misdirection 占 34.3%)、多文章声明比例(约 15.7%)以及平均词数(诚实完成 41.8 词,欺骗性完成 58.3 词)。已知限制包括长度差异、过滤偏差和欺骗风格不平衡。
FEVER Deception Source Dataset is a dataset of honest and deceptive fact-checking response pairs constructed based on the FEVER dataset, intended to provide training data for deception detection probes. Each sample includes a claim from FEVER with the ground-truth label SUPPORTS or REFUTES, an honest completion (which correctly derives the conclusion) and a deceptive completion (which deliberately employs seemingly plausible reasoning to reach an incorrect conclusion). The dataset is generated using the generative model `mlabonne/gemma-3-27b-it-abliterated` and the judging model `Qwen/Qwen3-30B-A3B`, and is built upon the SUPPORTS and REFUTES claims (excluding NEI) from the FEVER v1.0 training set. It consists of two batches (batch1_seed42 and batch2_seed43), with a total of 38,927 rows of data derived from 34,375 unique claims, and the claim duplication rate between batches is 23.4%. Statistical metrics include the distribution of claim types (each of SUPPORTS and REFUTES accounts for approximately 50%), deception styles (explicit_lie accounts for 65.7%, misdirection accounts for 34.3%), the proportion of multi-article claims (approximately 15.7%), and average word counts (honest completion: 41.8 words, deceptive completion: 58.3 words). Known limitations include length discrepancy, filtering bias, and imbalance in deception styles.




