Datasets:

CarolynJiang
/

RFC-Bench-Dataset

license: mit
task_categories:
  - text-classification
language:
  - en
tags:
  - Finance
  - Misinformation
pretty_name: BFC-Bench
size_categories:
  - 1K<n<10K

Dataset Description

RFC-Bench is a research benchmark for evaluating large language models on reference-free counterfactual financial misinformation detection at the paragraph level. The dataset contains only synthetic, human-reviewed counterfactual rewrites and metadata (ticker, date, and public URL) derived from publicly accessible Yahoo Finance articles; it does not redistribute any original news content.

All data concern corporate-level financial narratives and contain no personal or sensitive information. The dataset is released for academic research only and must not be used for investment advice, real-world financial communication, or commercial purposes.

The release includes a main benchmark (1,826 instances) and a hard-case subset (122 instances).

Citation

@article{jiang2026rfcbench,
  title   = {All That Glisters Is Not Gold: A Benchmark for Reference-Free Counterfactual Financial Misinformation Detection},
  author  = {Jiang, Yuechen and Liu, Zhiwei and Cao, Yupeng and He, Yueru and Xu, Ziyang and Xu, Chen and Deng, Zhiyang and Tiwari, Prayag and Chen, Xi and Lopez-Lira, Alejandro and Huang, Jimin and Tsujii, Junichi and Ananiadou, Sophia},
  journal = {arXiv},
  year    = {2026},
  eprint  = {2601.04160},
  url     = {https://arxiv.org/abs/2601.04160}
}