Datasets:

Languages:
English
Size Categories:
n>1T
ArXiv:
License:

Get values of `source` in dolma

#30
by maxmatical - opened

Is there a way to know what the value of source are beforehand, without downloading the entire dataset? For example gutenberg, common-crawl etc.

This would be useful if I want to download only a subset of the dataset by source, but don't want to download the entire dataset in order to get the set of unique sources.

Thank you :)

@maxmatical - help me understand your question better.
Are you asking for the set of all unique "source" values to be documented?
Or for a way to download the subset of dolma corresponding to a given "source" value of your choosing?

@dumitrac Ideally the first one. The main use would be if we want to try different sampling ratios in dolma (eg upsample flan, downsample refined web etc.)

If that is too difficult, would also be nice to be provided a list of source values, so I can process them on my end.

Thank you :)

Hi @maxmatical - I'm still confused about the question, please bear with me :D

The list of sources and their stats are documented in the dataset card:
image.png

Could you please give a sample file (json or text, etc) of what you're asking for my look like?

Hi @dumitrac . I meant more within the HF dataset object itself itself. For example from the dolma dataset, we have

{'id': 'a0d77923cd120285375f0afd739ecbc907ae990f',
 'text': 'The dawn of amateur radio in the U.K. and Greece: a personal view...',
 'added': '2023-04-25T05:49:48.253Z',
 'created': '1993-05-01T00:00:00.000Z',
 'source': 'gutenberg'}

What I would like to do is get all the unique values of source in order to split the dolma dataset into subsets by the various sources, eg if I only want the flan, gutenberg, and c4 subsets of dolma

Ideally we'd be able to do specify the subset of dolma we'd like to download (for example CohereForAI/aya_collection), but as long as I have the source values, I can work with that

Hi @dumitrac , just want to follow up with you to see if something like this is possible, thank you :)

Sign up or log in to comment