olmOCR-Bench

# olmOCR-Bench (pre-rendered pages) This dataset is a **pre-rendered page-image mirror** of [allenai/olmOCR-bench](https://huggingface.co/datasets/allenai/olmOCR-bench), prepared for use with [EvalScope](https://github.com/modelscope/evalscope). The original benchmark ships PDFs and requires every consumer to rasterize each page at evaluation time. To make evaluation fast, deterministic, and dependency-light, every referenced page has been rendered **once** to a PNG whose longest side is **2048 px**, matching the official renderer (`render_pdf_to_base64png`, `target_longest_image_dim=2048`). Consumers load the PNG directly and no longer need `pypdfium2` or the source PDFs. ## Contents - `bench_data/<source>.jsonl` — the official unit tests, copied verbatim from the upstream release (7 sources: `arxiv_math`, `headers_footers`, `long_tiny_text`, `multi_column`, `old_scans`, `old_scans_math`, `table_tests`). - `images/<pdf-path-without-.pdf>_p<page>.png` — one rendered page per `(pdf, page)` pair referenced by the unit tests (1,403 pages total). An image for a unit test is located as `images/{test['pdf'][:-4]}_p{test['page']}.png` (the `pdf` field keeps its source subfolder, e.g. `tables/...`). ## Provenance - Source data & unit tests: allenai/olmOCR-bench (ODC-BY-1.0), unchanged. - Page images: rendered from the source PDFs with `pypdfium2` at longest-side 2048 px, saved as PNG. - No unit test, target, or metadata was modified. ## License Distributed under **ODC-BY-1.0**, following the upstream dataset. See the [paper](https://arxiv.org/abs/2502.18443) and the [official evaluation code](https://github.com/allenai/olmocr) for details.