Datasets:

Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
ArXiv:
Libraries:
Datasets
Dask
License:

Difference with fineweb-edu-dedup (smol-lm corpus)

#106
by ferjorosa - opened

Hi, sorry if this has been asked before, but do you know the difference between this version of FineWeb-Edu and the one proposed by the Hugging Face team? (https://huggingface.co/datasets/HuggingFaceTB/smollm-corpus)

I am asking because both have been deduplicated but this one has 320M rows and the other has 190M

Airtrain AI org

Hi! So smollm-corpus is a corpus of multiple different datasets, with fineweb-edu being only a subset. In terms of their deduplicated fineweb-edu subset of smollm being different from ours--I don't know for certain but can speculate. Our duplicate detection was essentially as strict as possible; if a single character differed between two rows (ex: an extra space, capitalization differences, etc.) we kept both rows. There are a variety of ways you can loosen the criteria for what counts as a duplicate, and if you do so then you will treat more things as duplicates and thus have fewer rows in your deduplicated result. I am assuming they used looser criteria than we did.

Makes sense. Shame they do not comment more about it. Thanks, will investigate more!

Sign up or log in to comment