General

BIRD (text-to-SQL)

BIRD: a large-scale, cross-domain text-to-SQL benchmark (NL question + gold SQL over 95 real databases, 37 domains), scored by execution accuracy. The official repo ships questions/gold SQL/eval harness but no baseline predictions; this build ingests petavue/NL2SQL-Benchmark's public 'Inference Level Dataset' of 43,200 per-inference rows over a 360-question subset of BIRD dev. Each response is one model's binary execution-match correctness on one question under one prompting config; items are the real NL questions, correct_answer is the gold SQL. Models served under multiple provider aliases are merged to one canonical subject, with the serving platform recorded in test_condition.