FinanceBench overview and key facts (October 2026)

FinanceBench is a test set for question answering on company financial reports. Finance and data teams use the open sample to check whether a model or RAG setup finds correct figures in filings.

The open Hugging Face dataset contains 150 annotated examples: a question about a listed company, the answer and the evidence text from the source document. The PDFs are on GitHub. In the original paper, Patronus AI tested 16 model configurations, including vector stores and long context prompts, and reviewed 2,400 answers by hand. The full dataset is available from Patronus AI on request.

• Tests finance assistants before rolloutA finance team can run the 150 questions through its planned assistant and compare answers with the evidence. Errors such as wrong figures or invented numbers become visible before controllers rely on the tool.
• Compares retrieval setupsBecause each answer comes with its evidence string, teams can check whether their retrieval step found the right passage. This separates search errors from model errors. Teams know whether to fix the search index or the model prompt.
• Sets a minimum quality barThe questions are meant to be clear-cut. A model that fails them is not ready for more complex analysis work. Passing the test is a minimum requirement, and your own questions on company reports should follow before any rollout to finance staff.

• Each answer includes the evidence text from the source filing.
• Paper results include manual review of 2,400 model answers across 16 configurations.

FinanceBench uses questions about publicly traded companies based on their public filings; the linked PDFs are available on GitHub.

The open sample is licensed CC BY-NC 4.0, which excludes commercial use; clarify internal use with legal and ask Patronus AI about the full set. With 150 cases the sample is small, so add your own questions on your company's reports. The filings are US company documents in English.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews