pseudo-mini-pile / README.md
iNeil77's picture
Update README.md
78e6a7d verified
metadata
dataset_info:
  - config_name: all
    features:
      - name: content
        dtype: string
    splits:
      - name: train
        num_bytes: 360187653412.6177
        num_examples: 56194997
    download_size: 199030076349
    dataset_size: 360187653412.6177
  - config_name: c4_realnews
    features:
      - name: content
        dtype: string
    splits:
      - name: train
        num_bytes: 31597106256.723488
        num_examples: 11427438
    download_size: 19889880484
    dataset_size: 31597106256.723488
  - config_name: openwebtext
    features:
      - name: content
        dtype: string
    splits:
      - name: train
        num_bytes: 30974178275.039234
        num_examples: 6474479
    download_size: 19069709415
    dataset_size: 30974178275.039234
  - config_name: peS2o
    features:
      - name: content
        dtype: string
    splits:
      - name: train
        num_bytes: 221900508006.5479
        num_examples: 32612199
    download_size: 116217303065
    dataset_size: 221900508006.5479
  - config_name: redpajama_books
    features:
      - name: content
        dtype: string
    splits:
      - name: train
        num_bytes: 49246538575.26426
        num_examples: 107443
    download_size: 29612204926
    dataset_size: 49246538575.26426
  - config_name: stackexchange
    features:
      - name: content
        dtype: string
    splits:
      - name: train
        num_bytes: 2034535930.2150385
        num_examples: 716532
    download_size: 1222605537
    dataset_size: 2034535930.2150385
  - config_name: uspto
    features:
      - name: content
        dtype: string
    splits:
      - name: train
        num_bytes: 14755999149.910166
        num_examples: 3247716
    download_size: 7058272149
    dataset_size: 14755999149.910166
  - config_name: wiki
    features:
      - name: content
        dtype: string
    splits:
      - name: train
        num_bytes: 7528525537.163156
        num_examples: 1609190
    download_size: 4593971902
    dataset_size: 7528525537.163156
configs:
  - config_name: all
    data_files:
      - split: train
        path: all/train-*
  - config_name: c4_realnews
    data_files:
      - split: train
        path: c4_realnews/train-*
  - config_name: openwebtext
    data_files:
      - split: train
        path: openwebtext/train-*
  - config_name: peS2o
    data_files:
      - split: train
        path: peS2o/train-*
  - config_name: redpajama_books
    data_files:
      - split: train
        path: redpajama_books/train-*
  - config_name: stackexchange
    data_files:
      - split: train
        path: stackexchange/train-*
  - config_name: uspto
    data_files:
      - split: train
        path: uspto/train-*
  - config_name: wiki
    data_files:
      - split: train
        path: wiki/train-*
task_categories:
  - text-generation
language:
  - en
size_categories:
  - 10M<n<100M

A small, aggressively cleaned and de-duped pre-training corpus for academic settings. It aims to recreate something akin to The Pile but prioritizes quality for the constrained token budget academic researchers live with.

It has seven config subsets and an eighth all subset that combines them for a total of ~91B tokens (GPT2 Tokenizer estimate). These splits are as follows:

  1. c4_realnews: The RealNews domain subset of the C4 dataset containing news articles.
  2. openwebtext: The OpenWebText dataset containing the contents of the links mentioned in Reddit posts with at least 3 upvotes.
  3. peS2o: The PeS2o dataset containing academic articles from Semantic Scholar.
  4. redpajama_books: The books subset of RedPajama V1.
  5. stackexchange: The EN StackExchange non-code subset of the BigScience ROOTs dataset.
  6. uspto: The EN USPTO patent applications contents' subset of the BigScience ROOTs dataset.
  7. wiki: The EN Wiki subset of the BigScience ROOTs dataset.

The following processing and filtering steps have been applied:

  1. Removed citation text and bibliography information for academic texts.
  2. Ran a perplexity filter using a KenLM model trained on the EN OSCAR corpus and removed documents with a perplexity of more than 325 and less than 7.
  3. Removed samples which have a repeating <=4-gram proportion of 15%.
  4. Removed samples which have lower than 99% confidence of being EN using the lingua language detector.
  5. Performed an aggressive MinHash de-dupe using a shingle size of 8 and a low threshold of 0.5.