Datasets:

EleutherAI

/

pile

like 345

Update: It appears that the original Pile was taken down to avoid copyright infringement.
To avoid the same fate, I've posted pile-uncopyrighted, which is a backup of the entire Pile except for subsets with potentially infringing content.

Thanks so much for the solution! could you provide the information on which specific subset out of its original 22 sources are removed?

The information is in the "Methodology" section of the README. I'll put it here too:

Cleaning was performed by removing everything from the Books3, BookCorpus2, OpenSubtitles, YTSubtitles, and OWT2 subsets.
Based on section 7.1 of the original paper, these datasets are the only ones which are not explicitly allowed to be used in AI training.

Great!

Officially these file contents need update too:

_HOST_URL = "https://the-eye.eu"
_DATA_URLS = {
"all": {
"train": [f"{_HOST_URL}/public/AI/pile/train/{i:0>2}.jsonl.zst" for i in range(30)],
"validation": [f"{_HOST_URL}/public/AI/pile/val.jsonl.zst"],
"test": [f"{_HOST_URL}/public/AI/pile/test.jsonl.zst"],
},
"enron_emails": "https://www.cs.cmu.edu/~enron/enron_mail_20150507.tar.gz",
"europarl": f"{_HOST_URL}/public/AI/pile_preliminary_components/EuroParliamentProceedings_1996_2011.jsonl.zst",
"free_law": f"{_HOST_URL}/public/AI/pile_preliminary_components/FreeLaw_Opinions.jsonl.zst",
"hacker_news": f"{_HOST_URL}/public/AI/pile_preliminary_components/hn.tar.gz",
"nih_exporter": f"{_HOST_URL}/public/AI/pile_preliminary_components/NIH_ExPORTER_awarded_grant_text.jsonl.zst",
"pubmed": f"{_HOST_URL}/public/AI/pile_preliminary_components/PUBMED_title_abstracts_2019_baseline.jsonl.zst",
"pubmed_central": f"{_HOST_URL}/public/AI/pile_preliminary_components/PMC_extracts.tar.gz",
"ubuntu_irc": f"{_HOST_URL}/public/AI/pile_preliminary_components/ubuntu_irc_until_2020_9_1.jsonl.zst",
"uspto": f"{_HOST_URL}/public/AI/pile_preliminary_components/pile_uspto.tar",
"github": f"{_HOST_URL}/public/AI/pile_preliminary_components/github.tar",
}

can someone reupload?