Dedup granularity: are stack-v3-train repository rows complete, and do all files belong to the labeled repo?
#4
by iamgroot42 - opened
Thank you so much for the release! 🙌
The card says, under deduplication:
"For stack-v3-train, a single representative file per cluster is retained, chosen by (in order) highest stars, highest forks, permissive license, earliest repository creation."
I'm trying to understand how this interacts with the repository grouping, where each row is a full repository with code inline under files[].content
- When you say a representative file is kept per cluster, does "file" mean a single source file, or a whole repository? Since the MinHash clustering is described as language-agnostic and running across all repositories, I'm assuming it's individual files - can you confirm?
- Given file-level, cross-repo dedup, if I'm looking at a repository row in
stack-v3-train:
- (a) Is every row a complete copy of the repo, or can individual files be missing because their cluster's representative was retained under a different repository?
- (b) Are all files in a row guaranteed to originate from that exact repo, or could a retained "representative" file actually be the copy from an earlier fork / another repository that happened to win its cluster (higher stars/forks/earlier creation)?
For context on (b): since forks with ≥5 stars were pulled in and representatives are selected by stars/forks/creation date, I want to be sure a row labeled owner/repo contains that repo's own files, and not representative copies inherited from a fork.