BanglaDocBench

A document-parsing benchmark for Bangla: 500 to 1,000 pages chosen for diversity from 19.9 million.

Status: in development. No benchmark data has been released yet. This repository reserves the name and will hold the benchmark when it is finished.

BanglaDocBench will measure how well OCR and document-parsing systems read Bangla documents: books, newspapers, government notices, textbooks, forms, old prints and modern typesetting. It follows the approach of benchmarks such as OmniDocBench, for Bangla script.

How the pages are chosen

The pages are drawn from BanglaDocIndex, a public index of 19,873,418 Bangla document pages with an EmbeddingGemma 2 embedding for every page.

1. Clustering. The page embeddings are clustered to map the kinds of pages in the pool.
2. Selection. Pages are sampled across clusters for diversity, with a fixed random seed.
3. Annotation. Each selected page is annotated (text, reading order and layout).

The cluster labels, the selected page IDs, the seed and the selection code will be published with the benchmark, and the selected pages will be marked in BanglaDocIndex so that training sets can exclude them.

Licence

• Code: Apache-2.0, at github.com/OpenCR-lab/BanglaDocBench.
• Annotations: CC-BY-4.0 (planned).
• Page images are not relicensed. Each page image remains under the licence or terms of its source document. The source, the source URL and the source's own rights statement are recorded for every page. Many sources state no licence; copyright in those pages stays with their rights holders, and they are included for research and evaluation only. Anyone using the images is responsible for following the terms of each source.
• Removal requests: if you hold the rights to a page and want it removed, open an issue at github.com/OpenCR-lab/BanglaDocBench/issues with the page ID or source URL.