license: cc-by-4.0
pretty_name: T2-RAGBench
task_categories:

• table-question-answering
• question-answering
  language:
• en
  tags:
• pdf
• finance
• rag
  arxiv: 2506.12071
  configs:
• config_name: FinQA
  data_files:
  • split: train
    path: data/FinQA/train/*
  • split: dev
    path: data/FinQA/dev/*
  • split: test
    path: data/FinQA/test/*
• config_name: ConvFinQA
  data_files:
  • split: turn_0
    path: data/ConvFinQA/*
• config_name: TAT-DQA
  data_files:
  • split: train
    path: data/TAT-DQA/train/*
  • split: dev
    path: data/TAT-DQA/dev/*
  • split: test
    path: data/TAT-DQA/test/*

Dataset Card for T2-RAGBench

Project Page | Paper | Code

Table of Contents

• Dataset Card for T2-RAGBench
  • Table of Contents
  • Dataset Description
    • Dataset Summary
    • Supported Tasks
    • Leaderboards
    • PDF Files
    • Languages
  • Dataset Structure
    • Data Instances
    • Data Fields
      • FinQA and ConvFinQA Only
      • TAT-DQA Only
    • Data Splits
  • Dataset Creation
    • Curation Rationale
    • Source Data
    • Annotations
  • Personal and Sensitive Information
  • Considerations for Using the Data
    • Social Impact of Dataset
    • Discussion of Biases
    • Other Known Limitations
  • Additional Information
    • Licensing Information
    • Citation Information
    • Contributions

IMPORTANT NOTICE:

We deleted VQAonBD from the dataset due to low quality of the question reformulations. If you still want to use it you will find the data in the previous commit history.

Dataset Description

Dataset Summary

T2-RAGBench is a benchmark dataset designed to evaluate Retrieval-Augmented Generation (RAG) on financial documents containing both text and tables. It consists of 23,088 context-independent question-answer pairs and over 7300 documents derived from three curated datasets: FinQA, ConvFinQA, and TAT-DQA. Each instance includes a reformulated question, a verified answer, and its supporting context composed of textual and tabular information. It is also possible to use the pdfs directly, as the dataset includes the original PDF files.

Supported Tasks

• Question Answering (QA)
• Table-based Question Answering (TableQA)
• Retrieval-Augmented Generation (RAG)

Leaderboards

You can submit your results to the T2-RAGBench leaderboard to compare your model's performance against others.
The submission guidelines are available on the leaderboard page.

PDF Files

The dataset includes original PDF files from which the text and tables were extracted. These files can be used for direct document-based tasks or to verify the context of the questions. To download the PDF files clone the this repository and all files will be available in the data directory. The files are organized by dataset and split, matching the structure of the dataset.

Languages

• English

Dataset Structure

Data Instances

Each instance contains a unique identifier, a question, a context (text and table), and a verified answer.

Data Fields

For each subset, each sample contains the following fields:

• id: Unique identifier for the sample
• context_id: Identifier for the context document
• split: Dataset split (train, dev, test, turn_0, or validation_5)
• question: Context-independent QA query
• program_answer: Reformulated numeric answer used for evaluation
• original_answer: Original answer from the source dataset
• context: Extracted document text including both textual and tabular information
• file_name: Name of the source PDF file

FinQA and ConvFinQA Only

• table: Table content extracted from the PDF in Markdown format
• pre_text: Document text located before the table
• post_text: Document text located after the table
• company_name: Name of the company from the financial report
• company_symbol: Stock ticker symbol of the company
• report_year: Year of the financial report
• page_number: Page number in the PDF where the table was found
• company_sector: Sector classification of the company (e.g., Financials, Energy)
• company_industry: Industry classification of the company
• company_headquarters: Location of the company's headquarters
• company_date_added: Date the company was added to the reference index (e.g., S&P 500)
• company_cik: Central Index Key used by the SEC for company identification
• company_founded: Year the company was founded

TAT-DQA Only

• company_name: Name of the company from the financial report
• report_year: Year of the financial report
• company_sector: Sector classification of the company

Data Splits

Dataset Creation

Curation Rationale

Most existing QA datasets rely on oracle-contexts, which limit their ability to evaluate retrieval quality. T2-RAGBench transforms questions into a context-independent form to evaluate both retrieval and reasoning.

Source Data

Selected from existing QA datasets: FinQA, ConvFinQA, TAT-DQA. FinQA and ConvFinQA are based on FinTabNet. TAT-DQA is a diverse QA set with a focus on numerical answers.

Annotations

Questions were reformulated with LLaMA 3.3-70B to ensure context-independence. Human annotators verified a random subset of examples. Reformulated questions showed >80% context-independence compared to <10% in the originals.

Personal and Sensitive Information

Documents originate from public financial filings. No sensitive or personal user data is included. Entity names are real company names extracted from SEC filings.

Considerations for Using the Data

Social Impact of Dataset

T2-RAGBench encourages the development of RAG systems capable of reasoning over complex, real-world documents, such as those found in finance.

Discussion of Biases

The dataset focuses on financial documents, and domain-specific biases such as consistent formats or terminology may limit generalizability.

Other Known Limitations

• Reformulated questions are LLM-generated
• Performance evaluations may be influenced by prompt templates
• Dataset focused on finance domain only

Additional Information

We also have converted other SOTA benchmark in the same format.
They can be found in our collection:
https://huggingface.co/collections/G4KMU/rag-datasets

Licensing Information

CC-BY-4.0

Citation Information

@misc{strich2025t2ragbenchtextandtablebenchmarkevaluating,
      title={T$^2$-RAGBench: Text-and-Table Benchmark for Evaluating Retrieval-Augmented Generation}, 
      author={Jan Strich and Enes Kutay Isgorur and Maximilian Trescher and Chris Biemann and Martin Semmann},
      year={2025},
      eprint={2506.12071},
      archivePrefix={arXiv},
      primaryClass={cs.IR},
      url={https://arxiv.org/abs/2506.12071}, 
}

Contributions

This benchmark builds upon the following datasets:

• FinQA: Numerical reasoning over financial documents
• ConvFinQA: Conversational QA extension of FinQA
• TAT-DQA: Hybrid document QA with tables and text