davanstrien/newspapers-olmocr2

Document OCR using olmOCR-2-7B-1025-FP8

This dataset contains markdown-formatted OCR results from images in davanstrien/newspapers-with-images using olmOCR-2-7B.

Processing Details

• Source Dataset: davanstrien/newspapers-with-images
• Model: allenai/olmOCR-2-7B-1025-FP8
• Number of Samples: 100
• Processing Time: 0h 9m 29s
• Processing Date: 2025-10-22 18:57 UTC

Configuration

• Image Column: image
• Output Column: markdown
• Dataset Split: train
• Batch Size: 4
• Max Model Length: 16,384 tokens
• Max Output Tokens: 8,192
• GPU Memory Utilization: 80.0%

Model Information

olmOCR-2-7B is a high-quality document OCR model based on Qwen2.5-VL-7B-Instruct, fine-tuned on olmOCR-mix-1025 dataset and optimized with GRPO reinforcement learning.

Key features:

• 📐 LaTeX equations - Mathematical formulas in LaTeX format
• 📊 HTML tables - Structured table extraction
• 📝 Document structure - Headers, lists, formatting preserved
• 🖼️ Figure descriptions - Charts and figures labeled with descriptions
• 🔄 Rotation detection - Metadata about document orientation
• 📑 Natural reading order - Handles multi-column and complex layouts
• 🎯 High accuracy - Scores 82.4 ± 1.1 on olmOCR-Bench

Output Format

Each row contains:

• Original image from source dataset
• markdown: Extracted document content in markdown format
• olmocr_metadata: JSON with document metadata (language, rotation, table/diagram flags)

Columns

• image: Original document image
• markdown: Extracted text and structure in markdown
• olmocr_metadata: Document metadata (primary_language, is_rotation_valid, rotation_correction, is_table, is_diagram)
• inference_info: Processing metadata (model, script version, timestamp)

Reproduction

# Using HF Jobs (recommended)
hf jobs uv run --flavor l4x1 \
  -s HF_TOKEN \
  https://huggingface.co/datasets/uv-scripts/ocr/raw/main/olmocr2-vllm.py \
  davanstrien/newspapers-with-images \
  your-username/output-dataset

# Local with GPU
uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/olmocr2-vllm.py \
  davanstrien/newspapers-with-images \
  your-username/output-dataset

Citation

@misc{olmocr,
      title={{olmOCR: Unlocking Trillions of Tokens in PDFs with Vision Language Models}},
      author={Jake Poznanski and Jon Borchardt and Jason Dunkelberger and Regan Huff and Daniel Lin and Aman Rangapur and Christopher Wilhelm and Kyle Lo and Luca Soldaini},
      year={2025},
      eprint={2502.18443},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2502.18443},
}

---

Generated with uv-scripts/ocr