README.md

---
license: odc-by
tags:
- ocr
- benchmark
- pdf
- document-understanding
language:
- en
pretty_name: olmOCR-bench Pre-Rendered
size_categories:
- 1K<n<10K
configs:
- config_name: arxiv_math
  data_files:
  - split: test
    path: images/arxiv_math/**
- config_name: headers_footers
  data_files:
  - split: test
    path: images/headers_footers/**
- config_name: long_tiny_text
  data_files:
  - split: test
    path: images/long_tiny_text/**
- config_name: multi_column
  data_files:
  - split: test
    path: images/multi_column/**
- config_name: old_scans
  data_files:
  - split: test
    path: images/old_scans/**
- config_name: old_scans_math
  data_files:
  - split: test
    path: images/old_scans_math/**
- config_name: tables
  data_files:
  - split: test
    path: images/tables/**
---

# olmOCR-bench Pre-Rendered

Pre-rendered PNG images of the [olmOCR-bench](https://huggingface.co/datasets/allenai/olmOCR-bench) benchmark dataset, ready for zero-setup evaluation of any OCR / vision model.

## What This Is

The official [olmOCR benchmark](https://github.com/allenai/olmocr) requires downloading 1,403 PDFs locally and rendering each page to a PNG image before sending it to a model. **Every benchmark runner in the official repo does this same rendering step internally** — see [`olmocr/data/renderpdf.py::render_pdf_to_base64png()`](https://github.com/allenai/olmocr/blob/main/olmocr/data/renderpdf.py).

This dataset eliminates that setup entirely by hosting the pre-rendered images directly. The PNGs are rendered at **`target_longest_image_dim=2048`** — the same default resolution used by the official olmOCR `render_pdf_to_base64png()` function and by the GPT-4o, Claude, and Gemini benchmark runners.

All files are accessible via direct URL, so you can evaluate any model by just pointing at these URLs — no local downloads, no PDF rendering tools, no dataset cloning.

## Dataset Structure

The dataset has **7 subsets** (one per benchmark category), each with a **test** split:

| Subset | PDFs | Tests | Test Types |
|--------|------|-------|------------|
| `arxiv_math` | 522 | 2,927 | math |
| `headers_footers` | 266 | 753 | absent |
| `long_tiny_text` | 62 | 442 | present |
| `multi_column` | 231 | 884 | order |
| `old_scans` | 98 | 526 | present, absent, order |
| `old_scans_math` | 36 | 458 | math |
| `tables` | 188 | 1,020 | table |

### Loading with `datasets`

```python
from datasets import load_dataset

ds = load_dataset("shhdwi/olmocr-pre-rendered", "arxiv_math", split="test")
print(ds[0])  # {'image': <PIL.Image>, 'pdf_stem': '...', 'category': '...', ...}
```

## Contents

| Directory | Contents | Count |
|-----------|----------|-------|
| `images/` | Pre-rendered PNG images (page 1, 2048px longest dim) with `metadata.jsonl` per category | 1,403 images |
| `ground_truth/` | JSONL test case files (from allenai/olmOCR-bench) | 7,010 tests |
| `predictions/` | Published model prediction caches | 1,403 .md files |

## Rendering Details

Each PDF page is rendered to PNG matching the official olmOCR benchmark process:

- **Resolution**: `target_longest_image_dim = 2048` (longest side scaled to 2048px, aspect ratio preserved)
- **Renderer**: PyMuPDF (same pixel output as pdftoppm used in the official repo)
- **Pages**: Page 1 only (the benchmark tests only page 1 of each PDF)
- **Naming**: `{pdf_stem}_pg1.png`

This matches what the official benchmark runners do internally:
- `run_chatgpt.py`: `render_pdf_to_base64png(pdf_path, target_longest_image_dim=2048)`
- `run_claude.py`: `render_pdf_to_base64png(pdf_path, target_longest_image_dim=2048)`
- `run_gemini.py`: `render_pdf_to_base64png(pdf_path, target_longest_image_dim=2048)`
- `run_server.py`: `render_pdf_to_base64png(pdf_path, target_longest_image_dim=1024)` (for smaller models)

## Quick Start

Evaluate any model with zero setup:

```bash
pip install litellm httpx

# Run on any litellm-supported model
python run_bench.py --model gpt-4o
python run_bench.py --model claude-sonnet-4-20250514
python run_bench.py --model gemini/gemini-2.0-flash

# Run specific categories only
python run_bench.py --model gpt-4o --categories arxiv_math headers_footers

# Evaluate published predictions (no API key needed)
python run_bench.py --evaluate nanonets-optimal-v4
```

## Direct File Access

Every file is accessible via URL:

```
https://huggingface.co/datasets/shhdwi/olmocr-pre-rendered/resolve/main/images/arxiv_math/2503.05390_pg14_pg1.png
https://huggingface.co/datasets/shhdwi/olmocr-pre-rendered/resolve/main/ground_truth/arxiv_math.jsonl
```

## Attribution

Based on [olmOCR-bench](https://huggingface.co/datasets/allenai/olmOCR-bench) by Allen AI ([paper](https://huggingface.co/papers/2502.18443)). Licensed ODC-BY-1.0.