Datasets:

EleutherAI

/

pile-deduped-pythia-preshuffled

like 3

Hi there,
I have used this repo while following the instructions to set up the batch_viewer. I noticed that after a certain index (i.e., >=70_539_829), I only retrieve sequences with padding. That is

dataset = MMapIndexedDataset("path/to/merged/folder/document", skip_warmup=True)
dataset[70_539_829]  # array with all 0's

Can this be an issue with the unshading?

UPDATE: this has been solved by re-running the unsharding process