Synthetic Counterfactual Financial Paragraphs for Misinformation Detection

RFC-Bench is a collection of synthetic, human-reviewed counterfactual rewrites of financial news paragraphs for benchmarking large language models. The dataset was created by CarolynJiang and published on Hugging Face in January 2026. It includes metadata such as ticker, date, and source URL but does not contain original article text.

Use Cases

• Benchmarking LLM performance on reference-free detection of counterfactual financial misinformation using synthetic paragraph text.
• Analyzing the relationship between metadata fields like ticker and date and the presence of counterfactual claims.
• Training classifiers to identify synthetic counterfactual rewrites derived from public Yahoo Finance articles.

Strengths

• Human-reviewed synthetic counterfactual rewrites ensure label quality.
• Metadata includes specific fields like ticker, date, and source URL for context.

Limitations

• Dataset size, row count, and sample size are unknown.
• Contains only synthetic text, not original news content, limiting real-world distribution testing.
• Potential bias towards corporate financial narratives from Yahoo Finance sources.

Provenance

Source

Derived from publicly accessible Yahoo Finance articles.

Collection Method

Contains synthetic, human-reviewed counterfactual rewrites; original content is not redistributed.

Freshness

Last updated in January 2026.