Datasets:
File size: 4,658 Bytes
d847043 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 | """在公开的中文 Text-to-SQL 基准 CSpider 的 dev 集上评测，好和别人的数字比。
python scripts/eval_cspider.py --cspider cspider --model Qwen/Qwen3-4B-Instruct-2507 --adapter out/lora \
--out results/paper/cspider_main.json
--cspider 指向 CSpider 官方包解压出来的目录（taolusi/chisp 给的 Google Drive），里面要有 dev.json 和
database/<db_id>/<db_id>.sqlite。题目是中文，库是 Spider 的英文库，库里有真实数据。
和 eval.py 的不同：
- 库是现成的真数据，不用 seed_suite 生成，只在这一份库上比结果（Spider 官方的执行准确率也是这么算的）
- 建表语句和「部分字段的取值」从库里现读，和 ask.py 一样
- 主指标用严格口径（列数也要一样），和 Spider 的执行准确率一致；宽松口径照样记下来
repair / vote 和 eval.py 一样，看的是同一个库上的执行结果，不看标准答案。
"""
import argparse
import json
import sqlite3
from pathlib import Path
import eval as ev
from prompt import build, values_from_db
def open_db(path: Path) -> sqlite3.Connection:
conn = sqlite3.connect(f"file:{path}?mode=ro", uri=True)
conn.execute("PRAGMA query_only = ON")
# Spider 有几个库里混着不是 UTF-8 的字节，按默认的 str 读会直接抛错
conn.text_factory = lambda b: b.decode("utf-8", errors="replace")
return conn
def schema_of(conn: sqlite3.Connection) -> str:
return "\n\n".join(sql.strip() + ";" for (sql,) in conn.execute(
"select sql from sqlite_master where type='table' and name not like 'sqlite_%' and sql is not null"))
def main() -> None:
ap = argparse.ArgumentParser()
ap.add_argument("--cspider", required=True, help="CSpider 解压目录，含 dev.json 和 database/")
ap.add_argument("--model", required=True)
ap.add_argument("--adapter", default="")
ap.add_argument("--out", required=True)
ap.add_argument("--limit", type=int, default=0)
ap.add_argument("--max-new-tokens", type=int, default=1024)
ap.add_argument("--batch-size", type=int, default=8)
ap.add_argument("--no-values", action="store_true")
ap.add_argument("--repair", type=int, default=0)
ap.add_argument("--vote", type=int, default=1)
ap.add_argument("--timeout", type=float, default=30.0, help="单条 SQL 最多跑几秒；Spider 有几个库比较大")
args = ap.parse_args()
from infer import Generator, answer, execute
ev.QUERY_TIMEOUT = args.timeout
root = Path(args.cspider)
items = json.loads((root / "dev.json").read_text(encoding="utf-8"))
if args.limit:
items = items[: args.limit]
conns, ddls, values = {}, {}, {}
for db in sorted({it["db_id"] for it in items}):
conns[db] = open_db(root / "database" / db / f"{db}.sqlite")
ddls[db] = schema_of(conns[db])
values[db] = None if args.no_values else values_from_db(conns[db])
gen = Generator(args.model, adapter=args.adapter, batch_size=args.batch_size,
max_new_tokens=args.max_new_tokens)
convs = [build(ddls[it["db_id"]], it["question"], values[it["db_id"]]) for it in items]
outs = answer(gen, convs, [conns[it["db_id"]] for it in items], repair=args.repair, vote=args.vote)
records, ok, loose, valid, bad_gold = [], 0, 0, 0, 0
for it, o in zip(items, outs):
conn = conns[it["db_id"]]
try:
ev.rows_of(conn, it["query"])
runs, hit, hit_loose = ev.judge(it["query"], o["sql"], [conn])
except Exception: # noqa: BLE001
# 标准答案自己跑不通（超时之类）的题照样算在分母里、算错，记一下有几道
bad_gold += 1
runs, hit, hit_loose = execute(conn, o["sql"])[1] is None, False, False
valid += runs
ok += hit
loose += hit_loose
records.append({"db": it["db_id"], "question": it["question"], "gold": it["query"], "pred": o["sql"],
"hit": hit, "hit_loose": hit_loose, "runs": runs, "raw": o["raw"],
"first_pred": o["first_sql"], "votes": o["votes"], "repairs": o["repairs"]})
n = len(items)
print(f"\nCSpider dev 共 {n} 题｜可执行 {valid}（{valid/n:.1%}）｜执行准确率 {ok/n:.1%}"
f"｜宽松 {loose/n:.1%}｜标准答案跑不通 {bad_gold} 题")
tmp = Path(args.out + ".tmp")
tmp.parent.mkdir(parents=True, exist_ok=True)
tmp.write_text(json.dumps(records, ensure_ascii=False, indent=1), encoding="utf-8")
tmp.replace(args.out)
print("明细写到", args.out)
if __name__ == "__main__":
main()
|