sample split details
the original 1T dataset contained a certain % split from various sources: commonCrawl, arxiv, wikipedia etc..
does this 1B sample contain
- similar % split but with just less tokens, or
- its either heavily taken from one source, or
- randomly chosen 1B with no % of data source into account
thanks in advance :)
Hi @sujantkumarkv , we randomly sampled 0.1% of documents from each split to roughly match 1B tokens -- so the first option in your list applies.
@mauriceweber
When using this dataset withds = load_dataset("togethercomputer/RedPajama-Data-1T-sample")
, do we still need to sample in proportion to their source? Such asdata_config = [ ("arxiv", 2.5), ("book", 4.5), ("c4", 15.0), ("cc", 67.0), ("github", 4.5), ("stackexchange", 2.0), ("wikipedia", 4.5), ]
in here
What about the RedPajama-Data-1T-Sample dataset?
Hi @zgxiao , the sample dataset reflects the original proportions of the full dataset. So you don't need to pass custom weights for each source.