biglam/rubenstein-manuscript-catalog

Duke University Rubenstein Library Manuscript Card Catalog Dataset

Dataset Description

This dataset contains approximately 48,000 digitized catalog cards from the Duke University David M. Rubenstein Rare Book & Manuscript Library card catalog. The cards represent manuscript collections held by the library, documenting personal papers, organizational records, and historical manuscripts.

Why This Dataset?

Historical manuscript catalog cards are valuable sources of:

• Manuscript collection metadata in semi-structured formats
• Handwritten and typewritten text from multiple cataloging periods
• Historical archival practices and description methods
• Information about rare and unique primary source materials

This dataset is valuable for training and evaluating AI models on:

• OCR and handwriting recognition for archival documents
• Information extraction from catalog cards
• Named entity recognition (names, places, dates)
• Vision-language model performance on library/archival materials

Dataset Summary

• Total Cards: ~48,000 individual catalog cards
• Total Drawers: 69 drawers organized alphabetically
• Image Format: JPG (converted from JP2)
• OCR Coverage: ~100% (Tesseract 5.x via Internet Archive)
• Source: Internet Archive - Rubenstein Manuscript Catalog
• Original Institution: Duke University Rubenstein Library

Dataset Structure

Data Fields

Field	Type	Description
image	Image	JPG image of the catalog card (~800-1600px width)
drawer_id	string	Internet Archive item identifier (e.g., "rubensteinmanuscriptcatalog_A_to_Amer")
card_number	int	Card position within drawer (0-indexed)
filename	string	Original filename of the card image
text	string	OCR-extracted text from the card (Tesseract 5.x)
has_ocr	bool	Whether OCR text is available for this card
source	string	Attribution string
source_url	string	Direct link to source drawer on Internet Archive
ia_collection	string	Internet Archive collection identifier

Data Splits

This dataset contains a single train split with all catalog cards.

Example

from datasets import load_dataset

dataset = load_dataset("davanstrien/rubenstein-manuscript-catalog")

# Access a sample
sample = dataset['train'][0]
print(f"Card #{sample['card_number']}")
print(f"OCR Text:\n{sample['text']}")

# Display the card image
sample['image'].show()

Typical Card Contents

Manuscript catalog cards typically contain:

• Collection/Creator names (persons, families, organizations)
• Dates and date ranges (e.g., "1862", "1920-1945")
• Geographic locations (places of creation or subject)
• Subject descriptions (brief summaries of manuscript content)
• Physical extent (number of items, linear feet, volumes)
• Material types (letters, diaries, photographs, etc.)
• Call numbers and shelf locations
• Cross-references to related collections

Dataset Creation

Source Data

The cards were originally created by Duke University archivists and catalogers over many decades, documenting the library's manuscript holdings. The physical cards were digitized by Duke and uploaded to the Internet Archive with OCR processing.

Collection Process

1. Digitization: Duke University Rubenstein Library photographed each drawer of cards
2. OCR Processing: Internet Archive applied Tesseract OCR to create searchable text
3. Dataset Creation: Images were extracted from JP2 archives, converted to JPG, and paired with OCR text

Data Processing

• Original format: JP2 images in ZIP archives on Internet Archive
• Converted to: JPG at 95% quality for ML compatibility
• OCR format: HOCR (HTML with bounding boxes) → plain text extraction
• Organization: Preserved original drawer organization and card order

Uses

Example Use Cases

OCR Model Evaluation

from transformers import TrOCRProcessor, VisionEncoderDecoderModel

processor = TrOCRProcessor.from_pretrained('microsoft/trocr-base-handwritten')
model = VisionEncoderDecoderModel.from_pretrained('microsoft/trocr-base-handwritten')

# Evaluate on catalog cards with ground truth OCR
for sample in dataset['train'].select(range(100)):
    pixel_values = processor(sample['image'], return_tensors="pt").pixel_values
    generated_ids = model.generate(pixel_values)
    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

    # Compare with ground truth
    print(f"Ground truth: {sample['text'][:100]}")
    print(f"Generated: {generated_text[:100]}")

Named Entity Recognition

import re
from collections import Counter

# Extract dates from catalog cards
def extract_dates(text):
    # Pattern for years and date ranges
    pattern = r'\b(1[0-9]{3}|2[0-9]{3})(?:-([0-9]{2,4}))?\b'
    return re.findall(pattern, text)

# Analyze temporal distribution
dates = []
for card in dataset['train']:
    if card['text']:
        matches = extract_dates(card['text'])
        dates.extend(matches)

# Count most common decades
print(Counter(dates).most_common(10))

VLM Question Answering

from transformers import AutoProcessor, AutoModelForVision2Seq

processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-8b")
model = AutoModelForVision2Seq.from_pretrained("HuggingFaceM4/idefics2-8b")

# Ask questions about catalog cards
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "What is the name of the collection or creator on this card?"}
        ]
    }
]

for card in dataset['train'].select(range(10)):
    text = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(images=[card['image']], text=text, return_tensors="pt")
    outputs = model.generate(**inputs, max_new_tokens=100)
    print(processor.decode(outputs[0], skip_special_tokens=True))

Bias, Risks, and Limitations

Known Biases

1. Historical Language: Cards use terminology from various time periods that may be considered offensive or inaccurate by modern standards.

2. Collection Scope Bias: The collections reflect Duke University's historical collecting priorities and interests, which may not be representative of all manuscript repositories.

3. Cataloging Variation: Cataloging practices evolved over decades; older cards may use different description standards than newer ones.

OCR Limitations

• Handwriting Variation: Cards include handwritten entries with varying legibility
• Typewriter Quality: Older typewritten cards may have faded or inconsistent text
• Special Characters: Diacritical marks, symbols, or non-English text may be incorrectly recognized
• Layout Complexity: Annotations, cross-references, or struck-through text may confuse OCR

Additional Information

Dataset Curators

• Digitization: Duke University David M. Rubenstein Rare Book & Manuscript Library
• Hosting & OCR: Internet Archive
• Dataset Creation: Daniel van Strien (@davanstrien)

Licensing Information

The catalog card images are in the public domain. The dataset is released under CC0 1.0 Universal (Public Domain Dedication).

Citation Information

@dataset{rubenstein_catalog_2025,
  title={Duke University Rubenstein Library Manuscript Card Catalog Dataset},
  author={Duke University and van Strien, Daniel},
  year={2025},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/davanstrien/rubenstein-manuscript-catalog},
  note={Digitized catalog cards from Internet Archive collection rubensteinmanuscriptcatalog}
}

Acknowledgments

• Duke University Rubenstein Library for digitizing this invaluable archival resource and making it publicly accessible
• Internet Archive for hosting the collection and providing OCR processing
• The archivists and catalogers who created these cards over many decades, preserving knowledge about unique manuscript collections

Links

• Source Collection: https://archive.org/details/rubensteinmanuscriptcatalog
• Duke Rubenstein Library: https://library.duke.edu/rubenstein

Version History

• v1.0 (2025): Initial release with ~57,822 cards from 69 drawers

Contact

For questions about this dataset, please open an issue in the dataset repository