README.md

metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: is_filtered_out
      dtype: bool
  splits:
    - name: train
      num_bytes: 19005209693
      num_examples: 29451949
  download_size: 12244813118
  dataset_size: 19005209693
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Card for "wikipedia-bookscorpus-en-preprocessed"

Dataset Summary

A preprocessed and normalized combination of English Wikipedia and BookCorpus datasets, optimized for BERT pretraining. The dataset is chunked into segments of ~820 characters to accommodate typical transformer architectures.

Dataset Details

• Number of Examples: 29.4 million
• Download Size: 12.2 GB
• Dataset Size: 19.0 GB

Features:

{
    'text': string,       # The preprocessed text chunk
    'is_filtered_out': bool  # Filtering flag for data quality
}

Processing Pipeline

1. Language Filtering:

   • Retains only English language samples
   • Uses langdetect for language detection

2. Text Chunking:

   • Chunks of ~820 characters (targeting ~128 tokens)
   • Preserves sentence boundaries where possible
   • Splits on sentence endings (., !, ?) or spaces

3. Normalization:

   • Converts to lowercase
   • Removes accents and non-English characters
   • Filters out chunks < 200 characters
   • Removes special characters

4. Data Organization:

   • Shuffled for efficient training
   • Distributed across multiple JSONL files
   • No need for additional dataset.shuffle() during training

Usage

from datasets import load_dataset

dataset = load_dataset("shahrukhx01/wikipedia-bookscorpus-en-preprocessed")

Preprocessing Details

For detailed information about the preprocessing pipeline, see the preprocessing documentation.

Limitations

• Some tokens may be lost due to chunk truncation
• Very long sentences might be split
• Some contextual information across chunk boundaries is lost

Citation

If you use this dataset, please cite:

@misc{wikipedia-bookscorpus-en-preprocessed,
  author = {Shahrukh Khan},
  title = {Preprocessed Wikipedia and BookCorpus Dataset for Language Model Training},
  year = {2024},
  publisher = {GitHub & Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/shahrukhx01/wikipedia-bookscorpus-en-preprocessed}}
}