README.md

metadata

dataset_info:
  features:
    - name: text
      dtype: string
    - name: id
      dtype: string
    - name: dump
      dtype: string
    - name: url
      dtype: string
    - name: date
      dtype: string
    - name: file_path
      dtype: string
    - name: offset
      dtype: int64
    - name: token_count
      dtype: int64
    - name: language
      dtype: string
    - name: page_average_lid
      dtype: string
    - name: page_average_lid_score
      dtype: float64
    - name: full_doc_lid
      dtype: string
    - name: full_doc_lid_score
      dtype: float64
    - name: per_page_languages
      list: string
    - name: is_truncated
      dtype: bool
    - name: extractor
      dtype: string
    - name: page_ends
      list: int64
  splits:
    - name: train
      num_bytes: 4141357799
      num_examples: 186262
  download_size: 2040233575
  dataset_size: 4141357799
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Sampling Methodology

This dataset was created using reservoir sampling, a statistically unbiased random sampling algorithm that guarantees each sample from the source dataset has an equal probability of being included. This ensures the 1B token sample is representative of the full dataset's characteristics.

Source Dataset: HuggingFaceFW/finepdfs Sample Size: 1B tokens Content: High-quality textbook-style pdfs

Reservoir sampling enables rapid experimentation and ablation studies without processing the entire source dataset, while maintaining statistical validity of results.

For details on how this dataset was used in optimal pre-training data composition research, see the blog post.

Citation

If you use this model/dataset, please cite:

@article{sharma2025billion,
  title={The 1 Billion Token Challenge: Finding the Perfect Pre-training Mix},
  author={Sharma, Asankhaya},
  year={2025},
  url={https://huggingface.co/blog/codelion/optimal-dataset-mixing/}
}

For more details, see the blog post.