Datasets:
shuffling
#118
by
eminorhan
- opened
Thanks a lot for this dataset! Would it be possible to say a bit more about how the individual components of this dataset were shuffled (if at all)? While setting up distributed training, I'm using split_dataset_by_node to split the interleaved dataset into multiple dp (data-parallel) ranks. Would it be necessary/advisable to do some sort of shuffling (or pseudo-shuffling) before the split?
Hello! Thank you for using the dataset!
Yes, I would recommend to shuffle the shuffle the dataset, since we didn't shuffle it explicitly. It is also advisable to set the recommended weights for the components of the dataset.