77.1% on BIRD dev with one model call per question, and a test-set submission in

BIRD is the text-to-SQL benchmark the serious systems compete on: 1,534 dev questions over 11 real...