Organization of data
#2
by
jeffreygwang
- opened
I'm looking to make a random sample of a few chunks of The Pile (e.g. arxiv, pubmed, etc.). Since the official Pile mirror is down, I have stumbled upon this repository (thank you, @monology !).
I noticed that the samples are shuffled, but it's unclear in what way. How are they shuffled? If I were e.g. looking for a random sample of PubMed abstracts, could I just take the first .jsonl.zst file under "train" and grab all strings where 'pile_set_name' == 'PubMed Abstracts'?
The shuffle comes from the original Pile dataset, it was not done by me. At a glance, subsets appear to be evenly spread throughout all the chunks, so you'll most likely need to perform the filtering process on all the .jsonl.zst files.