pile-github / README.md
jqhoogland's picture
Update README.md
8fd1f81 verified
|
raw
history blame
1.01 kB
metadata
dataset_info:
  features:
    - name: text
      dtype: string
    - name: meta
      struct:
        - name: pile_set_name
          dtype: string
  splits:
    - name: train
      num_bytes: 542456309
      num_examples: 100000
  download_size: 202975613
  dataset_size: 542456309
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*

Dataset Creation Process

These subsets were created by streaming over the rows from the original dataset and filtering by the meta column. Each subset is generally limited to the first 100,000 qualifying rows encountered.

Note: Some subsets contain fewer than 100,000 rows due to the sparsity of certain categories in the original dataset. In these cases, the rows may not be the first encountered but are unevenly distributed due to parallelization during collection.

Future Updates

Future updates may include completing the smaller subsets to reach the 100,000 row target, which could result in breaking changes for those specific subsets.