You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

license: cc-by-nc-sa-4.0

Overview

Vision-language models perform well on single-document VQA, but their performance drops when they must retrieve the correct document from large, visually homogeneous collections. Invoice Haystack is designed to evaluate this challenging setting.

The benchmark contains:

1,500 anonymized invoice images
200 validated question-answer pairs
Three retrieval corpus scales: 500, 1,000, and 1,500 documents
Strong visual homogeneity, with a mean pairwise cosine similarity of 0.73

Downloads last month: -

Total file size:

5.48 GB