roberta_pretrain / README.md
gsgoncalves's picture
Update README.md
5b62716
|
raw
history blame
852 Bytes
metadata
license: unknown
task_categories:
  - fill-mask
  - text-generation
language:
  - en
pretty_name: RoBERTa Pretrain Dataset
size_categories:
  - 10M<n<100M

Dataset Card for RoBERTa Pretrain

Dataset Summary

This is the concatenation of the datasets used to Pretrain RoBERTa. The dataset is not shuffled and contains raw text. It is packaged for convenicence.

Essentially is the same as:

from datasets import load_dataset, concatenate_datasets
bookcorpus = load_dataset("bookcorpus", split="train")
openweb = load_dataset("openwebtext", split="train")
cc_news = load_dataset("cc_news", split="train")
cc_news = cc_news.remove_columns([col for col in cc_news.column_names if col != "text"])
cc_stories = load_dataset("spacemanidol/cc-stories", split="train")
return concatenate_datasets([bookcorpus, openweb, cc_news, cc_stories['train']])