Datasets:

ArmelR

/

the-pile-splitted

like 20

When I was processing ArXiv train, I found there are several small fragment of texts in the dataset:

'text': ['This\n',
  '= 10000\n',
  'paper.tex\n',
  '=1.8pc\n',
  '[99]{}\n',
  '0\n\n1\n',
  '\\\n',
  '',
  '= 10000\n',
  '=4800\n'],

The number of such fragments is 158 when I set the filtering condition to len(x['text']) <= 10.

So I'm wondering whether such a situation also appears in the original Pile-ArXiv, or only in the-pile-splitted-ArXiv? Does your splitting method potentially affect the completeness of the Pile? And, how can I deal with these fragments (if I have to filter them out, what would be the appropriate filtering condition)?