Datasets:
Question about dclm dedup
Hello,
I was comparing the dclm-dedup of zyda2 and found it has a different number of files from: https://huggingface.co/datasets/Zyphra/dclm-dedup
Is there a reason behind this difference?
This seems also reflected in a much larger download size. My current download of Zyda2 is already 5.9TB (in contrast with the 4.5TB indicated in the dataset sheet) and is still in the dclm dedup split, while the other version seems to only have 2TB?
Can you help me understand these differences?
+1 for this
Hi! Two versions are completely different. dclm-dedup is a full dedup of DCLM dataset. In Zyda-2 we did cross-dedup with fineweb-edu-2, so we removed much smaller amount of documents from DCLM.
Thank you for your answer. That clarifies my question.
And regarding the download size of Zyda-2, do you have an estimate for that value?
If you clone the repository, it'll be ~13TB total, here is a breakdown by every component: https://huggingface.co/datasets/Zyphra/Zyda-2#breakdown-by-component
Ah okay, I missed that column
Thanks for your reply!