T²-RAGBench
T²-RAGBench is a 2025 text-and-table benchmark for evaluating retrieval-augmented generation on financial documents, built by the Language Technology and HCDS groups at the University of Hamburg together with dida Datenschmiede GmbH. It assembles 23,088 context-independent question, context and answer triples over more than 7,300 real financial documents that mix prose with numerical tables. The questions come from three existing datasets, FinQA (8,281), ConvFinQA (3,458) and TAT-DQA (11,349), then reformulated with a large language model so each one names the entity, date and metric it refers to. The result is that every question maps to exactly one ground-truth document, which lets you score retrieval cleanly rather than handing the model its context for free.
You access it from Hugging Face, where it is auto-converted to Parquet and also readable as JSON, split into three configs that match the source datasets (a fourth, VQAonBD, was dropped for weak reformulations). Bulk download is straightforward and needs no credentials, and contexts are supplied in markdown so tables survive into your pipeline.
For RAG and AI work it is best used as an evaluation harness rather than a knowledge base. Because the questions are context-independent and numerical, it stresses both retrieval over table-heavy documents and the downstream numerical reasoning, and the authors report that hybrid sparse and dense retrieval works best here.
On licensing, the compilation is released under CC BY 4.0, so commercial use is permitted provided you attribute it. The caveat is the underlying material: the source datasets (FinQA, ConvFinQA, TAT-DQA) carry their own, mostly research-oriented terms, and the documents are public company filings with their own conditions, so check those before any commercial deployment. Compared with FinanceBench, which we also list and which is non-commercial and only partly open, T²-RAGBench is fully and openly downloadable and far larger, making it the more practical pick for iterating on financial, table-aware retrieval.
Related sources
CRAG (Comprehensive RAG Benchmark)
Meta's factual question answering benchmark of 4,409 QA pairs across five domains, paired with mock web-search and knowledge-graph retrieval APIs. It is distinct from "Corrective RAG (CRAG)", which is a retrieval method, not a dataset.
FinanceBench
A benchmark for evaluating LLMs and retrieval systems on questions about public financial filings. The full set holds 10,231 question, answer and evidence triplets; a 150 example sample is released openly.
FlashRAG Benchmark Datasets
A research toolkit that bundles standardised, pre-processed benchmark datasets for RAG evaluation across many scenarios. Every dataset comes in a single unified JSONL format, so you can swap benchmarks in and out without rewriting your data loading each time.
FRAMES (Fact, Fetch, and Reason)
A Google evaluation set of 824 challenging questions that tests factuality, multi-hop retrieval and reasoning together, each paired with a gold answer and the Wikipedia articles needed to answer it. Built to measure end-to-end RAG.