---

Dataset Card for FinBen-FOMC

Table of Contents

• Table of Contents
• Dataset Description
  • Dataset Summary
  • Supported Tasks and Leaderboards
  • Languages
• Dataset Structure
  • Data Instances
  • Data Fields
  • Data Splits
• Dataset Creation
  • Curation Rationale
  • Source Data
  • Annotations
  • Personal and Sensitive Information
• Considerations for Using the Data
  • Social Impact of Dataset
  • Discussion of Biases
  • Other Known Limitations
• Additional Information
  • Dataset Curators
  • Licensing Information
  • Citation Information
  • Contributions

Dataset Description

• Homepage: https://huggingface.co/datasets/TheFinAI/finben-fomc
• Repository: https://huggingface.co/datasets/TheFinAI/finben-fomc
• Paper: FinBen: An Holistic Financial Benchmark for Large Language Models
• Leaderboard: https://huggingface.co/spaces/finosfoundation/Open-Financial-LLM-Leaderboard

Dataset Summary

FinBen-FOMC is a financial sentiment classification dataset adapted from FOMC (Shah et al., 2023a). The dataset is designed for training and evaluating large language models (LLMs) on classifying central bank policy stances as Hawkish, Dovish, or Neutral.

Supported Tasks and Leaderboards

• Task: Hawkish-Dovish Classification
• Evaluation Metric: F1 Score, Accuracy
• Test Size: 496 instances

Languages

• English

Dataset Structure

Data Instances

Each instance consists of a structured format with the following fields:

• id: A unique identifier for each data instance.
• query: An excerpt from a central bank’s release.
• answer: The classification label (HAWKISH, DOVISH, or NEUTRAL).

Data Fields

• id: Unique string identifier for the data instance.
• query: The input text containing an excerpt from a central bank statement.
• answer: The classification label (HAWKISH, DOVISH, or NEUTRAL).

Data Splits

The dataset is split into:

• Test: 496 instances

Dataset Creation

Curation Rationale

The dataset is adapted from FOMC (Shah et al., 2023a) to improve its suitability for LLM-based classification tasks in central bank policy analysis.

Source Data

Initial Data Collection and Normalization

The dataset originates from Federal Open Market Committee (FOMC) statements and other central bank releases.

Who are the source language producers?

Central bank officials and policy documents.

Annotations

Annotation Process

Annotations follow a structured classification framework to label monetary policy stances.

Who are the annotators?

Financial experts and researchers.

Personal and Sensitive Information

No personally identifiable information (PII) is included.

Considerations for Using the Data

Social Impact of Dataset

This dataset enhances financial NLP capabilities, allowing more accurate analysis of monetary policy signals.

Discussion of Biases

Potential biases may exist due to:

• Interpretation differences in policy statements.
• Variability in central bank language across periods.

Other Known Limitations

• Requires financial domain expertise for best model performance.
• May not generalize well to non-FOMC policy documents.

Additional Information

Dataset Curators

• The Fin AI Team

Licensing Information

• License: CC BY-NC 4.0

Citation Information

Original Dataset:

@inproceedings{shah2023trillion,
  title={Trillion Dollar Words: A New Financial Dataset, Task & Market Analysis},
  author={Shah, Agam and Paturi, Suvan and Chava, Sudheer},
  booktitle={Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
  editor={Anna Rogers, Jordan Boyd-Graber, and Naoaki Okazaki},
  pages={6664--6679},
  year={2023},
  organization={Association for Computational Linguistics},
  address={Toronto, Canada},
  doi={10.18653/v1/2023.acl-long.368}
}

Adapted Version (FinBen-FOMC):

@article{xie2024finben,
  title={FinBen: A Holistic Financial Benchmark for Large Language Models},
  author={Xie, Qianqian and others},
  journal={arXiv preprint arXiv:2402.12659},
  year={2024}
}