T2-Ragbench
Benchmark financial QA combining text documents and tables to test RAG systems, with over 30,000 questions and original PDF documents.
32,908 QA pairs, 9000+ documents with text and table contexts, PDF and JSON
CC-BY 4.0
Description
T2-Ragbench is a benchmark designed to evaluate Augmented Generation by Recovery (RAG) models applied to financial documents. It combines complex documents — with text and tables — from sources like FinQA or ConvFinQA. Each sample includes a context-independent question, a verified answer, and a complete context derived from PDF documents. Original documentation is also provided for more advanced tasks.
What is this dataset for?
- Evaluate the ability of a model to extract relevant information from complex documents (text + tables)
- Testing the performance of RAG systems in a real financial context
- Building or refining specialized documentary QA systems
Can it be enriched or improved?
Yes. It is possible to add other types of financial documents, or to create multilingual variants. In addition, original PDF documents can be used to improve contextual annotations or to retrain specialized extractors.
🔎 In summary
🧠 Recommended for
- NLP engineers
- Documentary QA researchers
- RAG projects
🔧 Compatible tools
- Haystack
- LangChain
- Hugging Face Transformers
- PymuPDF
💡 Tip
Use original PDFs to train more robust contextual extractors or test parsing workflows.
Frequently Asked Questions
Can this dataset be used to train a complete RAG model?
Yes, it's specifically designed for that with rich QA pairs and a mixed text-table context extracted from documents.
Does this dataset only contain text or other types of data as well?
It contains both text and tables extracted from PDFs, making it a truly multi-modal data set.
Can we download the original PDFs?
Yes, PDFs are included and organized by folder. You only need to clone the linked repository to access it.
