How many documents are actually in Dolma v1.7?

#42
by yury-zyphra - opened

I downloaded the whole dataset and loaded it in huggingface dataset (after fixing a couple of issues with dolma.py), and I'm getting 3.4 billion documents instead of 2.5B claimed. Are there 0.9B extra duplicates?

I downloaded Dolma v1.6 as well, and my numbers matched what is claimed on the dataset card.

Same.

Here's what I got by downloading via a call to load_dataset. It took a bunch of reruns, but since the current version of datasets library attempts to resume downloads to the cache if a job dies or is killed in the middle, it eventually finished!

DatasetDict ({
    train: Dataset ({
        features: ['id', 'text', 'added', 'created', 'source'],
        num_rows: 3403336408
} )

Note that I also used the fix for duplicated ids in your prior comment https://huggingface.co/datasets/allenai/dolma/discussions/41.

Sign up or log in to comment