Dataset · pixparse
UCSF documents with OCR annotations
pixparse/idl-wds
Millions of document pages with PDF, TIFF images, and OCR annotations for recognition training.
What's inside
The dataset is filtered from the UCSF document library: 19 million pages deemed valid. Each document includes a PDF, a TIFF image with the same content, a JSON with extensive Textract OCR annotations, and an .ocr file with old OCR markup. Modalities are image and text; language is not specified in metadata.
What you can do with it
- Train an OCR model
- Fine-tune a vision-language model for image-to-text
- Evaluate text extraction from PDFs
- Tasks
- Image → text
- Size
- 10M–100M rows
- Downloads this month
- 16.9K
- Likes
- 207
Models trained on it
- HuggingFaceM4/idefics2-8bImage text → text 5.4K
- HuggingFaceM4/idefics2-8b-chattyImage text → text 538
- HuggingFaceM4/idefics2-8b-baseImage text → text 216
- HuggingFaceM4/idefics2-8b-AWQImage text → text 80
- mlx-community/idefics2-8b-4bitImage text → text 54
- mlx-community/idefics2-8b-chatty-4bitImage text → text 54
- HuggingFaceM4/idefics2-8b-base-AWQImage text → text 50
- HuggingFaceM4/idefics2-8b-chatty-AWQImage text → text 34
How to load it
from datasets import load_dataset
ds = load_dataset("pixparse/idl-wds")