Exact-source overlaps with stanford-star/dbinfer (4DBInfer): join-retailrocket is the same Kaggle RetailRocket dump as dbinfer-retailrocket; join-seznam (CTU) is table-for-table the same Seznam.cz wallet data as dbinfer-seznam. dbinfer-amazon (Books) and dbinfer-stackexchange (stats.SE) were already covered by the RelBench decontamination; AVS/diginetica/outbrain sources are not in the corpus. STATS: 641->639 DBs, 16,677->16,635 tasks. README decontamination section updated.
Drop join-amazon-home and join-se-askubuntu (preprocessed node blobs infeasible)
Their nodes.rkyv artifacts (~208 GB and ~132 GB) never made it into the-join-preprocessed and are not worth regenerating; both were already excluded from the curated rt-j.json training list. Removing the raw copies keeps the two repos consistent. STATS: 643->641 DBs, 16,791->16,677 tasks.
Decontaminate: drop 3 hidden stats.stackexchange copies (rel-stack source)
Removes join-stats, join-stats-ceb (CTU) and join-bird-codebase-community (BIRD): all three are the stats.stackexchange.com dump that rel-stack evaluates on, republished under non-SE names. Verified by identical SE schema and identical 1,032-row tags tables. Updates STATS catalogs (646->643 DBs, 16,882->16,791 tasks) and README. join-amazon-fashion retained: rel-amazon benchmarks books only.
Decontaminate: drop 4 DBs overlapping RelBench sources (rel-amazon/trial/ratebeer/f1)
overview.csv is not in the repo, so both links were dead, including the one used
for per-database license attribution. The catalogue lives in STATS/databases.parquet,
which the config header already exposes to the dataset viewer. Also links
STATS/tasks.parquet for the per-task catalogue.
The kind field was renamed from external to forecast in STATS/tasks.parquet and
in the per-task manifests, but the README still documented the old name. It also
claimed all tasks share one kind, which no longer holds now that autocomplete
tasks are present.
STATS: kind was 'external' for every row; take it from each task's manifest
timedelta was corrupted for 202 forecast tasks across 8 databases in two ways:
* 189 rows pinned at 1 nanosecond (degenerate fallback in the generator)
* 13 rows had the seconds count written into a nanosecond field
Corrected values are taken from each task's manifest.yaml, which was validated
against the real train timestamps (median gap between consecutive seed times)
for all affected tasks with zero relative error.
join-spider-bike-1 is deliberately unchanged: its 58.8 s window is genuine.
Also renames kind 'external' -> 'forecast' for the 6,220 forecast tasks, matching
db-task-lists/forecast.json. Autocomplete rows are untouched.
db-task-lists + STATS: switch autocomplete entries to explicit task names
