Neura Market
TAT-DQA logo

TAT-DQA

Free

a large-scale Document Visual Question Answering (VQA) dataset designed for complex document understanding, particularly in financial reports.

FreeFree tier
Type
Open Source
Visit Website GitHub (38 stars)

About TAT-DQA

TAT-DQA is a large-scale Document Visual Question Answering (VQA) dataset constructed by extending TAT-QA. It aims to stimulate progress in QA research over complex and realistic visually-rich documents that contain both tabular and textual content, especially those requiring numerical reasoning. The dataset is sampled from real-world high-quality financial reports, with an average of 550 words per document. It features diverse answer forms (single span, multiple spans, free-form) and various numerical reasoning capabilities (addition, subtraction, multiplication, division, counting, comparison, sorting, and their compositions). TAT-DQA contains 16,558 questions associated with 2,758 documents (3,067 document pages).

Key Features

Sampled from real-world high-quality financial reports with both tabular and textual data
Average document length of ~550 words, significantly larger than existing Document VQA datasets
Diverse answer forms: single span, multiple spans, free-form
Requires various numerical reasoning: addition, subtraction, multiplication, division, counting, comparison, sorting, and compositions
Contains 16,558 questions across 2,758 documents (3,067 pages)
Includes original PDF documents, OCR JSON output, and QA pairs with derivations

Pros & Cons

Pros
  • Large-scale annotated dataset with rich financial document content
  • Covers multiple numerical reasoning operations and answer types
  • Provides derivation steps for answers, enabling explainability evaluation
  • Open-source and freely available for download
Cons
  • Limited to the financial domain, reducing generalizability to other document types
  • Download requires Google Drive, no API or direct model access
  • Document language is English only (implied by examples)

Best For

Research in document visual question answering (VQA)Benchmarking numerical reasoning over mixed tabular and textual documentsDeveloping models for complex document understanding in financeEvaluating OCR and document layout parsing pipelines

FAQ

What is TAT-DQA?
TAT-DQA is a large-scale Document Visual Question Answering (VQA) dataset for complex document understanding, built by extending TAT-QA. It contains questions over real-world financial reports that require numerical reasoning.
How can I get started with TAT-DQA?
Download the dataset from Google Drive. It includes three parts: original PDF documents, converted content from PDF in JSON format (with bounding boxes), and QA pairs in JSON format.
What are the data formats?
Documents are in PDF format; OCR output is provided as JSON with pages, blocks, words, and bounding boxes; QA pairs are in JSON format with fields like question, answer, derivation, answer_type, and scale.
How many documents and questions are in TAT-DQA?
TAT-DQA contains 16,558 questions associated with 2,758 documents (3,067 document pages).