Datasets:

ArXiv:
Tags:

What's the concept on partitions

#5
by SwatCat - opened

Hi,

I have been trying to find what head, middle, tail portions mean. I went through CC website but found no relevant information. Can someone please explain.

Thanks

Together org

Hi @SwatCat ,

The head, middle and tail buckets partition the dataset into three parts based on the perplexity of a LM trained on Wikipedia. It is essentially an indicator of how similar the vocabulary (of a given document) is to the Wikipedia vocabulary. So, in general, you can expect the data in head and middle to be of higher quality compared to tail. However, this is only with respect to the Wikipedia model, so there will still be some good quality text in tail (just with a vocab different from Wikipedia) -- this is why we kept the tail bucket.

This partitioning comes from the CCNet pipeline which we have used to process the CommonCrawl data.

I hope this clarifies things!

Thank you, it does.

mauriceweber changed discussion status to closed

Sign up or log in to comment