metadata

task_categories:
  - text2text-generation
language:
  - en
pretty_name: Financial Context Dataset
size_categories:
  - 10K<n<100K

This dataset contains over 50,000 samples of user financial queries paired with their corresponding structured data requests (context). It was created to facilitate the creation of the Financial Agent LLM for accurate data extraction and query answering.

How to load the Dataset

You can load the dataset using the code below:

from datasets import load_dataset

ds = load_dataset("Chaitanya14/Financial_Context_Dataset")

Dataset Construction

Diverse Query Sources

The dataset was built using a multi-method query sampling approach, incorporating:

Real-world queries from retail customer technology channels (e.g., email, web forms).
Consultations with 10 financial advisors on common retail investor inquiries.
Interviews with 20 retail investors about their pre-investment questions.

Template-Based Expansion

Over 5,000 query templates were generated (4,000 from technology channels, 500 from advisors, and 500 from investors). These templates were then scaled up to over 50,000 samples by randomly varying elements within each template, ensuring a broad and varied representation of financial queries.

This dataset provides a valuable resource for developing and evaluating LLMs in the financial domain, particularly for applications requiring accurate understanding and response to user financial queries. For more information regarding the dataset, please refer to our paper.

Citation

If you use the Financial Context Dataset, please cite with the following BibTex entry:

@misc{sinha2025finbloomknowledgegroundinglarge,
      title={FinBloom: Knowledge Grounding Large Language Model with Real-time Financial Data}, 
      author={Ankur Sinha and Chaitanya Agarwal and Pekka Malo},
      year={2025},
      eprint={2502.18471},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2502.18471}, 
}