ilsilfverskiold/

ocr-benchmark

The 93 document images and ground truth used by the ocr-benchmark harness. The benchmark code, the reference run results, and the full methodology live in the GitHub repo — this dataset is the document corpus only.

One train split, 93 rows, one row per document:

Column Type Description

image Image The document page (PNG/JPG)

stem string Filename stem (e.g. invoice_000)

tier string Difficulty: easy, medium, or hard… See the full description on the dataset page: https://huggingface.co/datasets/ilsilfverskiold/ocr-benchmark.