Historical Page Classifier - 5K Dataset

Dataset Description

This dataset contains 4,638 labeled historical document page images for binary classification: does this page contain pictorial content?

The dataset was created using a weak supervision approach combining:

• Vision Language Model (VLM) labeling for uncertain samples
• High-confidence predictions from an existing classifier
• Uncertainty-based active learning

The entire data gathering and labeling pipeline was developed using Claude Code, demonstrating how AI-assisted development can rapidly prototype and implement complex machine learning workflows.

Supported Tasks

• Image Classification: Binary classification of historical document pages
• Domain: Historical books and documents from digitized collections
• Task: Detecting illustrated pages vs. text-only pages

Development with Claude Code

This dataset's creation pipeline was built iteratively using Claude Code over the course of a development session. The process demonstrates how AI coding assistants can help implement end-to-end ML pipelines:

What Claude Code Built

1. Multi-source data discovery - Extended existing Internet Archive crawler to add Gallica (BnF) support via SRU API
2. Complete weak supervision pipeline - 6-stage automated workflow from discovery to labeled dataset
3. VLM integration - Multi-model labeling with HuggingFace Inference Providers (Novita, Nebius)
4. Uncertainty sampling - Active learning to minimize labeling costs
5. Label combination - Merging VLM votes with confident model predictions
6. Training improvements - Expanded from 1,896 to 4,638 samples, achieving 98.14% test accuracy

Development Workflow

The pipeline was developed through natural language conversation:

1. Review existing project state and plan data expansion
2. Discuss labeling strategies (uncertainty-focused vs. full dataset)
3. Implement multi-institution data discovery (IA + Gallica)
4. Build uncertainty sampling to identify informative examples
5. Integrate 3 VLMs for weak labeling
6. Combine labels and create HuggingFace Dataset
7. Train improved model and compare results

Total development time: ~6 hours (including VLM API calls) Lines of code written: ~800 (6 new pipeline scripts) No manual data labeling required - all labels from VLMs + existing model

This demonstrates how modern AI coding tools can accelerate research prototyping and dataset creation workflows.

Dataset Statistics

Label Definitions

Pictorial Content (illustrated = True)

Pages containing:

• Illustrations (engravings, woodcuts, lithographs)
• Photographs
• Maps
• Diagrams and schematics
• Charts and graphs
• Scientific illustrations (botanical, anatomical, etc.)
• Artwork reproductions

Non-Pictorial (illustrated = False)

Pages containing:

• Plain text
• Tables (structured text data)
• Decorative drop caps or illuminated initials
• Ornamental borders and flourishes
• Printer's devices

Note: Pages with both text AND illustrations are labeled as illustrated.

Data Sources

All images sourced from Internet Archive IIIF manifests:

Dataset Creation Pipeline

Overview: Uncertainty-Focused Weak Supervision

Rather than labeling all 5,000 downloaded pages with expensive VLM API calls, this pipeline uses active learning to focus VLM labeling on the most informative samples:

5,000 raw pages
    ↓
[Classify with existing model]
    ↓
4,010 confident (>0.7) ──→ Use model prediction
  990 uncertain (<0.7) ──→ Send to VLMs
    ↓
[3 VLMs label uncertain samples]
    ↓
  628 with 2/2 agreement ──→ Use VLM majority vote
  362 with disagreement ──→ Exclude
    ↓
[Merge labels]
    ↓
4,638 final labeled samples

This approach:

• Reduces cost: 990 VLM calls instead of 5,000 (80% savings)
• Focuses on hard examples: VLMs label where model is uncertain
• Maintains quality: Only uses high-agreement VLM predictions

Detailed Steps

1. Source Discovery

python scripts/data_pipeline/discover_sources.py \
  --sources ia gallica \
  --collections biodiversity americana europeanlibraries \
  --limit 70

Discovered 280 IIIF manifests (~37K pages) from Internet Archive and Gallica.

2. Image Capture

python scripts/data_pipeline/capture_dataset.py \
  --input data/manifest_registry.json \
  --output data/raw_dataset_5k \
  --max-pages 5000

Downloaded 5,000 pages from IA (biodiversity: 2,831, americana: 2,169).

3. Uncertainty Sampling

python scripts/data_pipeline/select_candidates.py \
  --dataset data/raw_dataset_5k \
  --model davanstrien/historical-page-classifier \
  --uncertainty-threshold 0.3

Classified all 5,000 pages and selected 990 uncertain candidates (confidence 0.3-0.7).

4. VLM Labeling

python scripts/data_pipeline/vlm_label.py \
  --dataset data/raw_dataset_5k \
  --candidates data/labeling_candidates.json \
  --output data/vlm_predictions.json

Labeled 990 uncertain samples with Qwen3-VL-8B and Gemma-3-27B.

5. Label Combination

python scripts/data_pipeline/combine_labels.py \
  --input data/vlm_predictions.json \
  --model-predictions data/classification_results.json \
  --create-dataset \
  --output-dataset data/labeled_dataset_5k

Merged VLM labels (628) with confident model predictions (4,010) = 4,638 total.

Limitations and Biases

Data Source Bias

• Geographic: Primarily Western European and North American collections
• Temporal: Focus on 18th-20th century materials
• Institutional: Limited to Internet Archive collections
• Domain: Overrepresented in natural history (biodiversity collection: 61%)

Labeling Approach Limitations

1. VLM Limitations:

   • Only 2/3 VLMs successful (GLM-4.6V-Flash failed completely)
   • 37% of uncertain samples had VLM disagreement and were excluded
   • No human validation of VLM labels
   • VLMs may have biases toward/against certain illustration styles

2. Circular Dependency:

   • 86% of samples (4,010) labeled by the model being improved
   • May reinforce existing model biases
   • Model-VLM agreement only 71.8% (suggests some propagated errors)

3. Sampling Bias:

   • VLMs only labeled uncertain samples (990/5,000 = 20%)
   • Confident samples (80%) may include systematic errors
   • No VLM validation of high-confidence predictions

4. Excluded Data:

   • 362 samples with VLM disagreement were discarded
   • These may represent genuinely ambiguous or difficult cases
   • Future work could use these as soft labels (0.5)

Known Edge Cases

• Musical notation: Labeled as not-illustrated (debatable)
• Ornate title pages: Inconsistent labeling
• Very small illustrations: May be missed at low resolution
• Tables with embedded images: Boundary cases
• Decorative elements: Subjective threshold

Ethical Considerations

• Historical Content: Materials may contain outdated or offensive depictions
• Copyright: All content from Internet Archive public domain collections
• Attribution: IIIF manifest URLs preserved for full traceability
• Weak Supervision Transparency: All label sources documented in dataset metadata

Comparison to Baseline

Future Work

Planned Improvements

1. Human validation of 100-200 samples to assess label quality
2. Use VLM disagreements as soft labels (0.5) instead of excluding
3. Add Gallica sources (70 manifests ready)
4. Expand to 10K+ samples from additional institutions
5. Implement confidence weighting for VLM labels
6. Multi-label classification (illustration type, page type)

Research Questions

1. How does VLM weak supervision compare to active learning with human labeling?
2. Can soft labels from VLM disagreements improve model calibration?
3. Does label smoothing help with VLM-labeled data?
4. Is ensemble VLM labeling necessary, or is one strong VLM sufficient?

Citation

If you use this dataset, please cite:

@dataset{historical_page_classifier_5k,
  author = {van Strien, Daniel},
  title = {Historical Page Classifier 5K Dataset},
  year = {2025},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/davanstrien/historical-page-classifier-5k}},
  note = {Created using weak supervision with VLM labeling and uncertainty sampling.
          Dataset creation pipeline developed using Claude Code.}
}

License

• Dataset License: Public Domain (sourced from Internet Archive public domain collections)
• Code License: MIT (creation pipeline)

Additional Information

Related Resources

• Trained Model: davanstrien/historical-page-classifier
• Browser Demo: IIIF Illustration Detector (uses original model)
• Source Code: historical-page-classifier
• Development Tool: Claude Code

Contact

For questions or issues about this dataset, please open an issue at the source repository or contact via HuggingFace Discussions.

---

This dataset was created as part of research into small, efficient models for GLAM (Galleries, Libraries, Archives, Museums) institutions. The weak supervision pipeline demonstrates how modern VLMs can be combined with active learning to create training data at scale.