README.md

metadata

license: mit

KnowledgeMath Benchmark Description

KnowledgeMath is a knowledge-intensive dataset focused on mathematical reasoning within the domain of finance. It requires the model to comprehend specialized financial terminology and to interpret tabular data presented in the questions. KnowledgeMath includes 1200 QA examples across 7 key areas in finance. These examples were collected from financial experts and feature detailed solution annotations in Python format.

• Paper: https://arxiv.org/abs/2311.09797
• Code: https://github.com/yale-nlp/KnowledgeMath
• Leaderboard: will be released soon!

KnowledgeMath Dataset Information

All the data examples were divided into two subsets: validation and test.

• validation: 200 examples used for model development, validation, or for those with limited computing resources.
• test: 1000 examples for standard evaluation. We will not publicly release the annotated solution and answer for the test set.

You can download this dataset by the following command:

from datasets import load_dataset

dataset = load_dataset("yale-nlp/KnowledgeMath")

# print the first example on the validation set
print(dataset["validation"][0])

# print the first example on the test set
print(dataset["test"][0])

The dataset is provided in json format and contains the following attributes:

{
    "question_id": [string] The question id,
    "question": [string] The question text,
    "tables": [list] List of Markdown-format tables associated with the question, 
    "python_solution": [string] Python-format and executable solution by financial experts. The code is written in a clear and executable format, with well-named variables and a detailed explanation,
    "ground_truth": [integer] Executed result of `python solution`, rounded to three decimal places,
    "topic": [string] The related financial area of the question,
    "knowledge_terms": [list] List of knowledge terms in our constructed knowledge bank that is necessary to answer the given question. We will release this feature upon paper publication
}

Automated Evaluation

To automatically evaluate a model on KnowledgeMath, please refer to our GitHub repository here.

Citation

If you use the KnowledgeMath dataset in your work, please kindly cite the paper:

@misc{zhao2023knowledgemath,
      title={KnowledgeMath: Knowledge-Intensive Math Word Problem Solving in Finance Domains}, 
      author={Yilun Zhao and Hongjun Liu and Yitao Long and Rui Zhang and Chen Zhao and Arman Cohan},
      year={2023},
      eprint={2311.09797},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}