UDA FinQA — Question Answering (orgrctera/uda_fin_qa_qa)

Dataset description

This release is the FinQA-aligned Question Answering (QA) split from the UDA (Unstructured Document Analysis) benchmark: 8,190 expert-style question–answer instances grounded in real corporate financial disclosures (earnings materials, 10-K–style reports). Each row pairs a natural-language question with structured supervision (gold answers, supporting evidence, document context, and executable reasoning programs), so models can be trained or evaluated on financial numerical reasoning over heterogeneous evidence (narrative text and tables).

UDA (Hui et al., NeurIPS 2024) is a suite for Retrieval-Augmented Generation (RAG) and LLM-based analysis on real-world documents kept in original, messy formats. The finance track includes subsets such as FinHybrid (FinQA-style), designed to stress parsing, alignment, and reasoning—not only fluent generation.

FinQA (Chen et al., EMNLP 2021) is the foundational task: expert-written questions over financial reports with multi-step numerical reasoning, heterogeneous evidence (tables + text), and explainable annotations (including programs over quantities). This Hub dataset follows that task definition within UDA’s benchmark packaging (sub_benchmark: fin_qa).

The task

• Task type: Question Answering (QA) for FinQA — given the question in input, predict the correct numeric or textual answer using information that would appear in the source report (tables and surrounding text). Evaluation typically uses gold string and executable answers and may use evidence and program consistency with the official FinQA / UDA protocols.
• Input (input): A single English question about reported figures, trends, or relationships (e.g., interest expense, growth rates, year-over-year comparisons).
• Target (expected_output): A JSON string with:
  • answers: e.g. str_answer (normalized string) and exe_answer (numeric value where applicable).
  • evidence: Pointers to supporting text and table snippets (e.g. text_1, table_1).
  • context: Supporting pre_text, post_text, and table material aligned with the report excerpt.
  • program: A gold reasoning program over numbers and operations (FinQA-style), supporting interpretability and program-based metrics.
• Metadata (metadata): benchmark_name (uda_fin_qa), benchmark_type (uda), split, sub_benchmark (fin_qa), and value (JSON with identifiers such as label_key, label_file, q_uid).

Splits: Single split default with 8,190 examples (one Parquet shard: data/default-00000-of-00001.parquet).

Background

FinQA

FinQA was introduced to study numerical reasoning over financial data: questions are written by finance professionals over real filings; annotations include operations and facts that support the answer. The authors show that strong general-domain LMs still trail experts on finance-specific knowledge and multi-step numeric reasoning. FinQA remains a standard benchmark for table+text reasoning in finance.

UDA benchmark

UDA aggregates thousands of real documents and tens of thousands of annotated Q&A pairs across domains (including finance), with documents provided in ways that reflect realistic ingestion (e.g., PDF/HTML) so that retrieval, chunking, and parsing choices matter. The FinQA-related finance portion (FinHybrid / FinQA track) matches the scale of this dataset (8,190 QA instances).

References

FinQA (source task and annotations)

Zhiyu Chen, Wenhu Chen, Charese Smiley, Sameena Shah, Iana Borova, Dylan Langdon, Reema Moussa, Matt Beane, Ting-Hao Huang, Bryan Routledge, William Yang Wang. FinQA: A Dataset of Numerical Reasoning over Financial Data. EMNLP 2021, pages 3697–3711.

• Abstract (summary): The paper introduces a large-scale dataset of QA pairs over financial reports with gold reasoning programs and evidence for explainability; experiments show pretrained language models fall short of human experts on financial knowledge and complex numerical reasoning, motivating better structured and unstructured reasoning over filings.

• ACL Anthology: https://aclanthology.org/2021.emnlp-main.300/

• DOI: 10.18653/v1/2021.emnlp-main.300

• Original code and data: https://github.com/czyssrs/FinQA

• Project site: https://finqasite.github.io/

UDA (benchmark suite containing this FinQA slice)

Yulong Hui, Yao Lu, Huanchen Zhang. UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis. NeurIPS 2024 (Datasets and Benchmarks Track).

• Abstract (summary): UDA provides a benchmark for RAG in real-world document analysis with diverse domains and question types, using real documents and expert annotations; the work analyzes how parsing and retrieval interact with generation and highlights practical design choices for document AI systems.

• arXiv: https://arxiv.org/abs/2406.15187

• NeurIPS proceedings: https://proceedings.neurips.cc/paper_files/paper/2024/hash/7c06759d1a8567f087b02e8589454917-Abstract-Datasets_and_Benchmarks_Track.html

• Code: https://github.com/qinchuanhui/UDA-Benchmark

Related Hub resources

• Aggregated UDA QA reference: qinchuanhui/UDA-QA

Examples

Illustrative rows from the dataset; long context blocks are abbreviated.

Example 1 — interest expense

input:

what is the the interest expense in 2009?

expected_output (excerpt; context shortened):

{
  "answers": {
    "str_answer": "380",
    "exe_answer": 3.8
  },
  "evidence": {
    "text_1": "if libor changes by 100 basis points , our annual interest expense would change by $ 3.8 million ."
  },
  "context": {
    "pre_text": [
      "interest rate to a variable interest rate based on the three-month libor plus 2.05% ...",
      "if libor changes by 100 basis points , our annual interest expense would change by $ 3.8 million .",
      "..."
    ],
    "post_text": ["..."]
  },
  "program": "divide(100, 100), divide(3.8, #0)"
}

Example 2 — amortization growth

input:

what is the expected growth rate in amortization expense in 2010?

expected_output (excerpt):

{
  "answers": {
    "str_answer": "-27.0%",
    "exe_answer": -0.26689
  },
  "evidence": {
    "table_1": "fiscal years the 2010 of amortization expense is $ 5425 ;",
    "text_2": "amortization expense from continuing operations , related to intangibles was $ 7.4 million , $ 9.3 million and $ 9.2 million in fiscal 2009 , 2008 and 2007 , respectively ."
  },
  "context": { "...": "..." },
  "program": "subtract(1074.5, 1110.6), divide(#0, 1110.6)"
}

metadata.value (example):

{
  "label_key": "ADI_2009",
  "label_file": "fin_qa",
  "q_uid": "ADI/2009/page_49.pdf-1"
}

Citation

If you use this dataset, please cite FinQA, UDA, and this dataset record as appropriate:

@inproceedings{chen-etal-2021-finqa,
  title = {FinQA: 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 the 2021 Conference on Empirical Methods in Natural Language Processing},
  year = {2021},
  pages = {3697--3711}
}

@inproceedings{hui2024uda,
  title     = {UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis},
  author    = {Hui, Yulong and Lu, Yao and Zhang, Huanchen},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2024}
}

License

The original FinQA release is under the MIT License (see the FinQA repository). Use this dataset in compliance with the original data licenses and the UDA benchmark terms.