evaluation_set/check_gold.py

#!/usr/bin/env python3
"""Gold-label sanity checker for the GRAST-SL evaluation sets.

For every instance of every evaluation CSV in this folder it verifies:
  1. the gold column set is non-empty;
  2. every gold column is present in the instance's schema list
     (../schemas/<set>_schemas.json[db_id], exact case-insensitive match);
  3. n_schema_columns == len(schema list) for the instance's database.

The schema lists are flat `table.column` (Spider/BIRD) or grouped-node
(`FAMILY_*.column`, Spider 2.0) lists — the same granularity the gold
columns are written in, so membership is exact, no fuzzy matching.

Usage:  python check_gold.py            # check all four sets
        python check_gold.py bird_dev   # check one set
Exits non-zero if any check fails.
"""
import csv, json, os, sys

csv.field_size_limit(10_000_000)
HERE = os.path.dirname(os.path.abspath(__file__))
SCHEMAS = os.path.join(HERE, "..", "schemas")

SETS = {
    "spider_dev":       ("spider_dev.csv",       "spider_dev_schemas.json"),
    "bird_dev":         ("bird_dev.csv",         "bird_dev_schemas.json"),
    "spider2_lite_256": ("spider2_lite_256.csv", "spider2_lite_schemas.json"),
    "spider2_snow_256": ("spider2_snow_256.csv", "spider2_snow_schemas.json"),
}


def check(rows, schemas):
    bad = []
    for r in rows:
        rid = r.get("instance_id") or r.get("question_id")
        db = r["db_id"]
        schema = schemas.get(db)
        if schema is None:
            bad.append((rid, db, "NO_SCHEMA_FOR_DB", ""))
            continue
        up = {s.upper() for s in schema}
        cols = [c for c in r["gold_columns"].split("; ") if c]
        if not cols:
            bad.append((rid, db, "EMPTY_GOLD", ""))
            continue
        miss = [c for c in cols if c.upper() not in up]
        if miss:
            bad.append((rid, db, "NOT_IN_SCHEMA", "|".join(miss)))
        if "n_schema_columns" in r and int(r["n_schema_columns"]) != len(schema):
            bad.append((rid, db, "N_SCHEMA_MISMATCH",
                        f"csv={r['n_schema_columns']} len(schema)={len(schema)}"))
    return bad


def main():
    targets = sys.argv[1:] or list(SETS)
    failed = False
    for name in targets:
        csv_file, schema_file = SETS[name]
        rows = list(csv.DictReader(open(os.path.join(HERE, csv_file))))
        schemas = json.load(open(os.path.join(SCHEMAS, schema_file)))
        bad = check(rows, schemas)
        n_cols = sum(len([c for c in r["gold_columns"].split("; ") if c]) for r in rows)
        status = "OK" if not bad else f"{len(bad)} PROBLEM(S)"
        print(f"{name}: {len(rows)} instances, {n_cols} gold columns -> {status}")
        for b in bad[:20]:
            print("   ", *b)
        failed |= bool(bad)
    sys.exit(1 if failed else 0)


if __name__ == "__main__":
    main()