UDA TAT-QA (orgrctera/uda_tat_qa)

Overview

This dataset is the TAT-QA slice of the UDA (Unstructured Document Analysis) benchmark: 14,703 question–answer instances derived from real financial reports, packaged for retrieval-oriented evaluation in RAG pipelines.

UDA is a benchmark suite for Retrieval-Augmented Generation (RAG) over messy, real-world documents (PDF/HTML) where evidence mixes narrative text and tables. In the finance track, UDA includes TatHybrid—TAT-QA–aligned labeling at the scale of this release (170 documents, 14,703 Q&A pairs in the UDA paper’s finance configuration).

TAT-QA (Zhu et al., ACL 2021) is a large-scale QA benchmark over hybrid contexts: each instance combines semi-structured tables with multiple paragraphs from the same financial report. Questions require diverse numerical reasoning (e.g. arithmetic, counting, comparison) and may be answered with spans, numbers, or derived values. UDA adopts TAT-QA–style supervision within its broader document-analysis benchmark (Hui et al., NeurIPS 2024 Datasets & Benchmarks).

In this Hub release, each row is a retrieval task instance: systems must retrieve the right document regions (tables and text) and ground answers in that evidence—consistent with the UDA setting, where parsing, chunking, and retrieval are first-class concerns alongside generation.

Task

• Task type: Retrieval (within a RAG / document-analysis pipeline) for TAT-QA-style financial QA over table + text hybrid reports.
• Input: A natural-language question (input) about figures, policies, or relationships disclosed in corporate financial materials.
• Supervision / reference: expected_output is a JSON string with gold answers, supporting facts, and optional derivation traces (see below). metadata records UDA identifiers (sub_benchmark: tat_qa).

Evaluation typically combines retrieval quality (whether the correct passages or cells are retrieved) with answer correctness (span match, numeric accuracy, or reasoning alignment), following TAT-QA and UDA protocols.

Background

TAT-QA

TAT-QA targets question answering over hybrid tabular and textual content from real financial reports. Compared with text-only or table-only QA, it stresses joint reasoning: models must align numbers in tables with narrative explanations, perform multi-step operations, and handle scale and answer typing (e.g. span vs. arithmetic). The original paper reports strong human performance relative to early neural baselines, underscoring dataset difficulty.

UDA benchmark

UDA revisits RAG and LLM-based document analysis across domains using 2,965 real-world documents and 29,590 expert-annotated Q&A pairs, with sources kept in original formats to stress parsing and retrieval as well as generation. The TatHybrid finance subset corresponds to this dataset’s 14,703 examples in the default split.

Data fields

Column	Type	Description
input	string	Question text posed over the report.
expected_output	string	JSON string with fields such as answers (answer, answer_type, scale), facts (gold evidence strings), and derivation (reasoning trace when applicable).
metadata	struct	benchmark_name (uda_tat_qa), benchmark_type (uda), split, sub_benchmark (tat_qa), and value (JSON string with identifiers like label_key, label_file, q_uid, doc_page_uid).

Splits: Single split default with 14,703 examples.

Examples

The following rows are taken from the dataset (JSON in expected_output is shown formatted for readability).

Example 1 — span answer (benefits)

• input: What benefits are provided by the company to qualifying domestic retirees and their eligible dependents?
• expected_output:

{
  "answers": {
    "answer": ["certain postretirement health care and life insurance benefits"],
    "answer_type": "span",
    "scale": ""
  },
  "facts": ["certain postretirement health care and life insurance benefits"],
  "derivation": ""
}

Example 2 — arithmetic answer (pension interest cost)

• input: What is the change in Interest cost on benefit obligation for pension benefits from December 31, 2018 and 2019?
• expected_output:

{
  "answers": {
    "answer": 129,
    "answer_type": "arithmetic",
    "scale": ""
  },
  "facts": ["1,673", "1,802"],
  "derivation": "1,802-1,673"
}

metadata.value (structure, example):

{
  "label_key": "overseas-shipholding-group-inc_2019",
  "label_file": "tat_qa",
  "q_uid": "bbdcf6da614f34fdb63995661c81613f",
  "doc_page_uid": "2ef48dc98e756493f097d01acf8101a2"
}

References

TAT-QA (source task & data lineage)

Fengbin Zhu, Wenqiang Lei, Youcheng Huang, Chao Wang, Shuo Zhang, Jiancheng Lv, Fuli Feng, Tat-Seng Chua. TAT-QA: A Question Answering Benchmark on a Hybrid of Tabular and Textual Content in Finance. ACL 2021, pages 3086–3101.

• Abstract (short): Introduces TAT-QA from real financial reports with hybrid table-and-text contexts; requires numerical reasoning (e.g. arithmetic, counting, comparison). Proposes TAGOP and shows a large gap to human performance.
• ACL Anthology: https://aclanthology.org/2021.acl-long.254/
• arXiv: https://arxiv.org/abs/2105.07624
• Project page: https://nextplusplus.github.io/TAT-QA/
• Original data (reference): next-tat/TAT-QA on Hugging Face

UDA benchmark (suite containing this TAT-QA 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 (short): Presents UDA with thousands of real-world documents and tens of thousands of expert-annotated Q&A pairs; evaluates LLM- and RAG-based document analysis and highlights parsing and retrieval design choices.
• 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 & resources: https://github.com/qinchuanhui/UDA-Benchmark

Related Hub resources

• UDA QA aggregation (reference): qinchuanhui/UDA-QA

Citation

If you use this dataset, please cite both TAT-QA and UDA (and this dataset record as appropriate):

@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 Huang, Youcheng and Wang, Chao and Zhang, Shuo and Lv, Jiancheng and Feng, Fuli and Chua, Tat-Seng},
  booktitle = {Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)},
  year = {2021},
  pages = {3086--3101}
}

@article{hui2024uda,
  title   = {UDA: A Benchmark Suite for Retrieval Augmented Generation in Real-world Document Analysis},
  author  = {Hui, Yulong and Lu, Yao and Zhang, Huanchen},
  journal = {arXiv preprint arXiv:2406.15187},
  year    = {2024}
}

License

Use this dataset in compliance with the original TAT-QA and UDA data licenses and terms. The TAT-QA project page and repository document licensing for the underlying benchmark; verify conditions for your use case before redistribution or commercial use.