Light of History: Historical Event Classification

This repository contains code for training and deploying a binary classifier to identify historical events in Wikipedia articles. The project uses DistilBERT to classify text as either "historical event" or "not historical event," achieving validation accuracy above 99.5%.

Project Components

• Training pipeline using DistilBERT for binary classification
• Wikipedia article processing and classification system
• Validation and testing frameworks
• Batch processing capabilities for large-scale classification

Technical Details

• Model: DistilBERT with binary classification head
• Training Data: 211,525 balanced examples
• Split: 80/20 train/validation
• Hardware: Google Colab with NVIDIA T4 GPU
• Implementation: TensorFlow/Hugging Face Transformers

Resources

• Training dataset available on HuggingFace: guscastilloa/dbpedia_historical_events
• Model weights and classification outputs archived on Zenodo: doi:10.5281/zenodo.14126493

Key Features

• High-accuracy binary classification (>99.5% validation accuracy)
• Efficient processing of large text datasets
• Probability scores for classification decisions
• Edge case handling for ambiguous historical events

Future Development

• Implementation of k-fold and time-based cross-validation
• Scaling to full Wikipedia dataset (9M+ articles)
• Enhanced error analysis
• Integration of human-validated data

Citation

If you use this code in your research, please cite our Zenodo archive:

@dataset{castillo_2024_14126493,
  author       = {Castillo Alvarez, G. A. and
                  Garcia-Figal, A. and
                  Lage-Castellanos, A. and
                  Baez-Silva, A. M.},
  title        = {{Light of History: Model weights and classified output}},
  month        = nov,
  year         = 2024,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.14126493},
  url          = {https://doi.org/10.5281/zenodo.14126493}
}