Datasets:

EleutherAI

/

pile-deduped-pythia-preshuffled

like 3

Hi there,

I am trying to understand how the tokenized Pile dataset has been "packed" into sequences of fixed length (2049 tokens). I started wondering about this when noticing (from a small sample) that the sequences are not concatenated using the <EOS> token, which is something I though was done by default.

Thanks a lot for your help!