Official dataset repository for our LREC 2026 paper: BankMathBench: A Benchmark for Numerical Reasoning in Banking Scenarios

BankMathBench is a benchmark for evaluating numerical reasoning in realistic banking scenarios.
This repository provides train/test datasets across three difficulty levels:

• 1_basic
• 2_intermediate
• 3_advanced

Each level includes English (eng) and Korean (kor) versions.

All counts below are the number of data rows (header excluded).

CSV files contain banking scenario questions and numeric answers.

• Common columns: index, question, reasoning, answer

Unless explicitly stated otherwise in a question:

• Installment savings (적금) uses formulas based on deposits made at the beginning of each period and maturity at the end of each period.
• Interest is calculated on a pre-tax basis.

If you use this dataset, please cite:

@article{bankmathbench2026,
  title={BankMathBench: A Benchmark for Numerical Reasoning in Banking Scenarios},
  journal={arXiv preprint arXiv:2602.17072},
  year={2026}
}

본 데이터셋 및 저장소 내 자료의 저작권은 주식회사 카카오뱅크에 있습니다.
Copyright (c) KakaoBank Corp. All rights reserved.