FinRAG

Dataset Description

This dataset contains 12,500 financial reasoning questions based on real-world financial documents, earnings reports, and financial tables. Each question is accompanied by a correct answer and four carefully crafted distractor answers, making it suitable for multiple-choice question answering tasks and assessing financial numerical reasoning capabilities.

Dataset Summary

• Total Examples: 12,500
• Format: Multiple-choice questions with 5 options (1 correct + 4 distractors)
• Domain: Financial documents, earnings reports, financial tables
• Task: Numerical reasoning over financial text and tables
• Language: English

Source Data

This dataset combines all splits (train, validation, and test) from:

• FinQA: Financial Question Answering dataset (7,750 questions, 62%)
• TAT-QA: Table-and-Text Question Answering dataset (4,750 questions, 38%)

Distractor Generation

Four distractor answers were algorithmically generated for each question using the following techniques:

• Stop early: Stopping calculation before completion
• Negate operand: Negating numbers in the calculation
• Operand bleeding: Using the wrong operands from the table
• Replace operator: Using the wrong mathematical operation (e.g., multiply instead of add)
• Switch order: Changing the order of operations
• Percentage error: Mistakes in percentage conversion
• Unit error: Mistakes in unit conversion (e.g., millions vs. thousands)
• Append operation: Adding extra unnecessary operations
• Substitution error: Substituting incorrect values from the table

These techniques create plausible but incorrect answers that test true understanding of the financial reasoning task.

Dataset Structure

Data Fields

Each example in the dataset contains:

• id (string): Unique identifier for each question
• pre_text (list of strings): Contextual text passages from the financial document that appear before the table
• post_text (list of strings): Additional contextual text passages that appear after the table (may be empty)
• table (list of lists): Financial table data in row-major format, where the first row typically contains headers
• question (string): The financial reasoning question to be answered
• choices (list of strings): List of 5 answer choices (1 correct + 4 distractors), randomly shuffled
• answer (integer): Index (0-4) pointing to the correct choice in the choices list

Data Example

{
  "id": "9bbb9fb3-3482-4d4d-be40-dd6ff47c23e9",
  "pre_text": [
    "Orders at Mobility grew to a record high on a sharp increase in volume...",
    "Revenue grew slightly as double-digit growth in the customer services business..."
  ],
  "post_text": [],
  "table": [
    ["", "", "Fiscal year", "", "% Change"],
    ["(in millions of €)", "2019", "2018", "Actual", "Comp."],
    ["Orders", "12,894", "11,025", "17 %", "16 %"],
    ["Revenue", "8,916", "8,821", "1 %", "0 %"]
  ],
  "question": "Analyse this data from a financial earnings document. What it the increase / (decrease) in revenue from 2018 to 2019?",
  "choices": ["-3978", "17737", "94", "95", "1"],
  "answer": 3
}

Dataset Sources and Attribution

This derived dataset is based on two source datasets that are licensed under Creative Commons Attribution 4.0 International (CC-BY 4.0). The distractor generation methodology and dataset structure are licensed under MIT. This work complies with the CC-BY 4.0 attribution requirements by providing proper attribution, copyright notices, license information, and links to the original datasets below.

FinQA Dataset

Citation:

@inproceedings{chen-etal-2021-finqa,
    title = "{F}in{QA}: A Dataset of Numerical Reasoning over Financial Data",
    author = "Chen, Zhiyu  and Chen, Wenhu  and Smiley, Charese  and Shah, Sameena  and
              Borova, Iana  and Langdon, Dylan  and Moussa, Reema  and Beane, Matt  and
              Huang, Ting-Hao  and Routledge, Bryan  and Wang, William Yang",
    booktitle = "Proceedings of EMNLP 2021",
    year = "2021"
}

License: Creative Commons Attribution 4.0 International (CC-BY 4.0) Original Dataset: https://finqasite.github.io/ Modifications: This derived work adds multiple-choice distractors to the original FinQA questions.

TAT-QA Dataset

Citation:

@inproceedings{zhu-etal-2021-tat,
    title = "{TAT}-{QA}: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance",
    author = "Zhu, Fengbin  and Lei, Wenqiang  and Wang, Chao  and Zheng, Jianming  and
              Poria, Soujanya  and Chua, Tat-Seng",
    booktitle = "Proceedings of ACL-IJCNLP 2021",
    year = "2021"
}

License: Creative Commons Attribution 4.0 International (CC-BY 4.0) Original Dataset: https://nextplusplus.github.io/TAT-QA/ Modifications: This derived work adds multiple-choice distractors to the original TAT-QA questions.