Datasets & licensing

gr.datasets.load("ml-1m") fetches + caches + parses. Every entry records a license, citation, and download policy:

Policy	Datasets	Behavior
auto	MovieLens 100K/1M/25M/32M/latest/latest-small	auto-fetched for your own use (GroupLens permits research download)
auto_nc	KGRec, Last.fm (Taste Profile)	non-commercial — needs load(..., accept_license=True)
manual	CAMRa2011, Mafengwo, Weeplaces, Yelp, Douban	prints where to download + where to drop files

Licensing: the library is MIT (that's the code); datasets keep their own licenses. These never conflict because we never redistribute the data — auto fetches from the canonical host onto your machine; anything non-redistributable is manual. Note that fetching for your own use is not redistribution: the MovieLens terms even differ by release — the older 100K/1M READMEs forbid redistribution without permission, while 25M/32M/latest permit it under the same terms (the registry records the right one per dataset). ml-latest and ml-latest-small are rolling releases, so we pin each one's exact snapshot by sha256 for reproducibility.

ml-latest / ml-latest-small are development datasets

GroupLens states both are development datasets that "may change over time and [are] not an appropriate dataset for shared research results". Use ml-25m / ml-32m for anything you report; the registry records this in each spec's notes.

Group-structured datasets

gd = gr.datasets.load_consrec("path/to/CAMRa2011")   # AGREE/ConsRec format
# -> GroupBenchmarkData(dataset, groups, group_interactions, test_instances)

Use gd.test_instances with evaluate_sampled / benchmark(level="sampled").

Bring your own

gr.datasets.from_path("ratings.csv", user_col="u", item_col="i", rating_col="r")
gr.datasets.from_huggingface("some/repo", user_col="userId", item_col="movieId")

Preprocessing

data = gr.datasets.k_core(data, k=5)
data = gr.datasets.filter_min_interactions(data, min_per_user=5, min_per_item=5)
data = gr.datasets.binarize(data, threshold=4.0)