rasinmuhammed/retail-sql-evaluation-rotated

retail_answer_key

An evaluation database generated from its answer key. Every expected answer was declared before the data existed, then confirmed by running the gold SQL with DuckDB against these exact files. 46 questions over 3 databases (seeds 42 to 44), one split each. No source data was used.

The databases share every declared answer and differ in every row. A query is correct only when it returns the expected answer on all of them, so a query that fits particular rows (a hard-coded key, a LIMIT without ORDER BY) fails.

Of 123 plausible wrong queries (the gold SQL with one mistake), 11.4% return the expected answer on a single database and 11.4% on all 3 (manifest.json, near_miss_study).

Tables

Table	Rows
stores	40
customers	3,000
sales	40,000

Files

• questions.jsonl: one question per line with its gold SQL and expected answer
• certificate.json: the per-question check, per database
• manifest.json: the full spec, seeds, versions and dropped candidates
• rotations/<seed>/tables/: the database as CSV, pinned by attestation.intoto.json (156 of 156 invariants hold)

Score a model

pip install duckdb
python verify.py                             # every gold query, every database
python verify.py --candidates answers.jsonl  # your queries: {"id": ..., "sql": ...} per line

Generated with Misata 0.9.8.