peS2o-final / README.md
iNeil77's picture
Update README.md
20a8b23 verified
metadata
dataset_info:
  features:
    - name: content
      dtype: string
  splits:
    - name: train
      num_bytes: 134025361660
      num_examples: 19450435
  download_size: 68402599157
  dataset_size: 134025361660
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
size_categories:
  - 10M<n<100M

This is a filtered subset of the allenai/peS2o dataset which has been:

  1. Stripped of references using a combination of regexes.
  2. Aggressively minhash deduped using a shingle size of 20 and a threshold of 0.5.
  3. Run through a Lingua English language filter with a confidence threshold of 0.99.
  4. Run through an OSCAR trained English KenLM model (at edugp/kenlm) perplexity filter for values between 8 and 550.