bokane/hucker-bakeoff2-ovis-hard

Document OCR using OvisOCR2

This dataset contains OCR results from images in bokane/hucker-ocr-bakeoff-hard using OvisOCR2, a compact 0.9B document parsing model (96.58 on OmniDocBench v1.6).

Processing Details

• —Source Dataset: bokane/hucker-ocr-bakeoff-hard
• —Model: ATH-MaaS/OvisOCR2
• —Number of Samples: 8
• —Processing Time: 4.5 min
• —Processing Date: 2026-08-13 22:28 UTC

Configuration

• —Image Column: image
• —Dataset Split: train
• —Batch Size: 16
• —Max Model Length: 32,768 tokens
• —Max Output Tokens: 16,384
• —Temperature: 0.0 (greedy, per model card)
• —GPU Memory Utilization: 80.0%
• —Visual-region image tags: filtered (default)

Model Information

OvisOCR2 is a compact, high-performance document parsing model:

• —0.9B parameters (post-trained from Qwen3.5-0.8B with SFT + RL + OPD)
• —96.58 on OmniDocBench v1.6 (first end-to-end model to top the leaderboard)
• —Markdown output in natural reading order
• —LaTeX formula recognition, HTML table extraction
• —Apache-2.0 licensed

Dataset Structure

The dataset contains all original columns plus:

• —markdown: The extracted text in markdown format
• —inference_info: JSON list tracking all OCR models applied to this dataset

Reproduction

Produced on Hugging Face Jobs (gpu) with the `ovis-ocr2.py` recipe from uv-scripts. Run it yourself:

hf jobs uv run https://huggingface.co/datasets/uv-scripts/ocr/raw/main/ovis-ocr2.py \
    bokane/hucker-ocr-bakeoff-hard \
    <output-dataset> \
    --image-column image \
    --batch-size 16