Five of the eleven dev databases differ between dev.zip and minidev.zip

[ivermin1123]

The two zips the README links carry the same eleven dev_databases, but only six of the files are the same. I checked https://bird-bench.oss-cn-beijing.aliyuncs.com/dev.zip (sha256 cdd6d19faeb45a23970b98d3ef6c40a87987c95459c2cf12076897a60cf5a630, Last-Modified 2024-06-29) against https://bird-bench.oss-cn-beijing.aliyuncs.com/minidev.zip (sha256 cc48ba16838204e4e214512030cb572eeb5f7bcdd999bae4b9b6ff12ec13b92f, downloaded 2026-09-02) by file digest first, then by schema, row count per table and a digest of each table's rows where the files differed. card_games, codebase_community, debit_card_specializing, financial, student_club and superhero are byte-identical. The other five are not, and neither README says so.

Database	Table	dev.zip	minidev.zip
formula_1	lapTimes	420,369 rows	400,524 rows
formula_1	pitStops	6,070	5,815
formula_1	qualifying	7,397	6,967
formula_1	races	976	954
formula_1	results	23,657	23,179
thrombosis_prediction	Examination	806	106
toxicology	atom	12,333	9,111
toxicology	bond	12,379	9,156
toxicology	connected	24,758	18,312
california_schools	satscores	same count, 211 of 2,269 cds codes without their leading zero	the same codes with the zero
european_football_2	Player	same count, height stored as REAL	height stored as a truncated integer

The schemas agree in all five. In formula_1, thrombosis_prediction and toxicology the dev copy simply holds more rows. In california_schools the row counts agree and 211 of the 2,269 satscores.cds values lost their leading zero in dev's copy, so those 211 rows join to schools in Mini-Dev's copy and do not in dev's. In european_football_2 the counts agree and Player.height is a REAL in dev and a truncated integer in Mini-Dev.

Why it matters: a Mini-Dev question is a dev question, and its gold runs against whichever copy the user unpacked. A gold over formula_1, toxicology or thrombosis_prediction can return other rows on the two copies, and a gold that joins satscores to schools returns other rows on california_schools. Two people scoring the same predictions against "the dev databases" can get different EX for that reason alone, and the README of neither repository says which copy is the one the golds were written against.

Two questions, then: which copy is the intended one for Mini-Dev, and could the README of each zip state the sha256 of the other and the five differences? The check itself is a short script; I am happy to send it, or the per-table row digests, if that helps.