Blame for scripts/selftest.py

"""开 GPU 之前的自检：不用显卡、不联网，几分钟跑完。哪一项没过就别花钱开机器。

    python scripts/selftest.py

查这些：
1. 四个测试集：标准 SQL 能跑、结果不空、没有并列导致的结果不唯一、没有重复题（check_testset.py）
2. 评分本身：标准答案当模型输出判对；故意改坏的 SQL 判错；多返回几列宽松判对、严格判错；
   ROUND 过的数字、'2024' 和 2024 判对；抽 SQL 能处理代码块、CTE、前后的废话
3. 训练集能从 data/pairs + data/new 原样重建出来（数据、prompt、取值提示一个字都不差）
4. 训练集里没有测试库的题、没有跑不通的 SQL
5. 装了 sqlglot 的话：测试集标准答案没有「一对多连接后重复求和」
"""

import json
import re
import sqlite3
import sys
import tempfile
from pathlib import Path

import build_sft
from check_testset import check
from eval import ROOT, extract_sql, judge, load_dbs, same_result, superset_result
from seed_suite import make_suite

# 已知、可以接受的问题：价格最贵的课在生成的数据里经常没人约，结果是 0，题是对的，只是区分度低
KNOWN = {("gym", "#17 价格最贵的那个课程被预约了几次？")}

failed = []


def ok(name: str, cond: bool, detail: str = "") -> None:
    print(("  通过  " if cond else "  失败  ") + name + (f"（{detail}）" if detail and not cond else ""))
    if not cond:
        failed.append(name)


def main() -> None:
    print("1. 测试集体检")
    for t in ("library", "gym", "hard", "transit"):
        items = [json.loads(l) for l in (ROOT / f"data/test_{t}.jsonl").read_text(encoding="utf-8").splitlines() if l.strip()]
        problems, _ = check(items)
        real = [p for p in problems if not any(t == k and p.startswith(q) for k, q in KNOWN)]
        ok(f"test_{t}（{len(items)} 题）", not real, "; ".join(p.splitlines()[0] for p in real[:3]))

    print("2. 评分")
    cases = {
        "SELECT a FROM t;": "SELECT a FROM t",
        "```sql\nWITH x AS (SELECT 1) SELECT * FROM x;\n```": "WITH x AS (SELECT 1) SELECT * FROM x",
        "SQL：SELECT a FROM t; -- 说明": "SELECT a FROM t",
        "我不会": "",
    }
    ok("抽 SQL", all(extract_sql(src) == want for src, want in cases.items()))
    ok("ROUND 过的数字判对", same_result([("A", 0.28125)], [("A", 0.2813)], False))
    ok("'2024' 和 2024 判对", same_result([("2024", 3)], [(2024, 3)], False))
    ok("差得多的数字判错", not same_result([("A", 0.28125)], [("A", 0.29)], False))
    ok("多返回一列：严格错、宽松对",
       not same_result([("x",)], [(1, "x")], False) and superset_result([("x",)], [(1, "x")], False))
    ok("少返回一列：宽松也错", not superset_result([("x", 1)], [("x",)], False))
    for t in ("library", "gym", "hard", "transit"):
        items = [json.loads(l) for l in (ROOT / f"data/test_{t}.jsonl").read_text(encoding="utf-8").splitlines() if l.strip()]
        db = items[0]["db"]
        _, conns = load_dbs(db, make_suite(db, [i["sql"] for i in items], 5))
        gold_ok = sum(judge(i["sql"], i["sql"], conns)[1] for i in items)
        ok(f"test_{t} 标准答案当输出全判对", gold_ok == len(items), f"{gold_ok}/{len(items)}")
        # 把第一个 WHERE / HAVING 条件换成永假，绝大多数应该判错
        broken = [re.sub(r"(?i)\b(WHERE|HAVING)\b", r"\1 1 = 0 AND", i["sql"], count=1) for i in items]
        wrong = sum(judge(i["sql"], b, conns)[2] for i, b in zip(items, broken) if b != i["sql"])
        ok(f"test_{t} 改坏的 SQL 基本判错", wrong <= 2, f"{wrong} 条改坏的被判对")

    print("3. 训练集能原样重建")
    with tempfile.TemporaryDirectory() as tmp:
        out = Path(tmp) / "train.jsonl"
        sys.argv = ["build_sft.py", "--out", str(out)]
        build_sft.main()
        same = out.read_bytes() == (ROOT / "data/train.jsonl").read_bytes()
    ok("build_sft 重建结果和 data/train.jsonl 一字不差", same,
       "数据或脚本改过，跑一遍 python scripts/build_sft.py")

    print("4. 训练集干净")
    rows = [json.loads(l) for l in (ROOT / "data/train.jsonl").read_text(encoding="utf-8").splitlines() if l.strip()]
    ok("没有测试库的题", not any(r["db"] in build_sft.TEST_DBS for r in rows))
    broken = 0
    for db in {r["db"] for r in rows}:
        conn = sqlite3.connect(":memory:")
        conn.executescript(build_sft.schema_path(db).read_text(encoding="utf-8"))
        for r in rows:
            if r["db"] == db:
                try:
                    conn.execute(r["messages"][-1]["content"].rstrip(";")).fetchall()
                except sqlite3.Error:
                    broken += 1
    ok(f"{len(rows)} 条 SQL 都能在建表语句上跑", broken == 0, f"{broken} 条跑不通")

    print("5. 重复求和（需要 sqlglot）")
    try:
        from lint_fanout import fanout_problems
    except ImportError:
        print("  跳过  没装 sqlglot（pip install sqlglot）")
    else:
        flagged = 0
        for t in ("library", "gym", "hard", "transit"):
            for l in (ROOT / f"data/test_{t}.jsonl").read_text(encoding="utf-8").splitlines():
                if l.strip():
                    it = json.loads(l)
                    flagged += bool(fanout_problems(it["sql"], it["db"]))
        ok("测试集标准答案没有重复求和", flagged == 0, f"{flagged} 条")

    print("\n全部通过" if not failed else f"\n{len(failed)} 项没过：{failed}")
    sys.exit(1 if failed else 0)


if __name__ == "__main__":
    main()