Datasets:

Size Categories:
n>1T
ArXiv:
License:

Dataloader Separation on Clean & Noisy Subset

#6
by sabilmakbar - opened

Hello, I'd like to suggest modifying the current dataloader configuration for HF Dataloader. I've noticed an issue where obtaining the clean subset requires downloading its corresponding noisy counterpart. This is because the dataloader currently treats language as config choices, with ("clean", "noisy") as its subset. Unfortunately, this design can be discouraging as it consumes more space (both clean and noisy subsets must be loaded), leading to longer overall processing times during the load_dataset operation.

To address this, I propose the following adjustment:
The configuration should be separated for languages with clean and noisy subsets. For example, "abt" could be divided into "abt_noisy" and "abt_clean." This change would result in each split containing only one train subset, improving efficiency and minimizing unnecessary resources. From the current implementation, I think the changes can be minimal and I'd be more than happy to help on this.

Sign up or log in to comment