Dataset · hashmortar
Table, PDF, and document question answering
hashmortar/spreadsheet-bench-v2-modified
A set of 1,060 questions and reference answers in English across 171 business documents in various formats.
What's inside
The dataset contains 1,060 English question-reference answer pairs tied to 127 Excel tables, 35 PDF files, and 9 DOCX documents. This is a revised version of the SpreadsheetBench 2 benchmark, shifting the focus from table editing to finding, interpreting, and synthesizing facts from multiple sources. All source files are taken from open sources, and the annotations are newly written.
What you can do with it
- Evaluating RAG system quality on mixed business documents (Excel, PDF, DOCX)
- Testing multimodal and tabular models in Multi-Document QA tasks
- Benchmarking AI agents for corporate reporting analysis
- Tasks
- Question answering, Table question answering, Document question answering
- Languages
- English
- Size
- 1K–10K rows
- Downloads this month
- 22.8K
- Likes
- 2
How to load it
from datasets import load_dataset
ds = load_dataset("hashmortar/spreadsheet-bench-v2-modified")