Meta Kaggle Submissions: stream the 2.4 GB CSV straight to parquet

A Hugging Face mirror of a dated Meta Kaggle snapshot serves Competitions.csv (124 MB), Teams.csv (727 MB) and Submissions.csv (2.4 GB) as plain resolve/main URLs, so you don't need a Kaggle API token. On a nearly full disk, pipe curl -sSL .../Submissions.csv into pandas read_csv(sys.stdin, usecols=[Id, TeamId, IsAfterDeadline, IsSelected, PublicScoreFullPrecision, PrivateScoreFullPrecision], chunksize=2_000_000), drop rows missing either score, and write each chunk with a pyarrow ParquetWriter. The result is 16.5M rows in about 350 MB, built in under a minute. Gotchas: IsSelected has mixed types across chunks (DtypeWarning); DeadlineDate is '%m/%d/%Y %H:%M:%S'; EvaluationAlgorithmAbbreviation is an opaque numeric id for many 2023+ competitions; and simulation or forecasting competitions have public == private or a constant private score, so a public-private gap scale of exactly 0 is a quick way to filter them out.