Datasets:
Download scripts/teacher.py from YukiKaze1337/cn-sqlite-sft: direct link, hf CLI and curl.
- Browser
- Download file 6.1 kB
-
https://huggingface.co/datasets/YukiKaze1337/cn-sqlite-sft/resolve/main/scripts/teacher.py
- Command line
-
hf download hf://datasets/YukiKaze1337/cn-sqlite-sft/scripts/teacher.py
-
curl -L -o teacher.py https://huggingface.co/datasets/YukiKaze1337/cn-sqlite-sft/resolve/main/scripts/teacher.py
6.1 kB
| """让老师模型照着建表语句造 (中文问题, SQLite 查询) 对。 | |
| python scripts/teacher.py --schema schemas/school.sql --n 20 | |
| 生成的 SQL 会真的丢进内存 SQLite 跑一遍(数据用 seed_suite 现生成),跑不过的直接扔掉; | |
| 只留单条 SELECT / WITH,DDL/DML 一律不要。结果追加到 data/new/<库名>.jsonl, | |
| 跑完用 verify.py 复核,再用 build_sft.py 拼进训练集。 | |
| """ | |
| import argparse | |
| import json | |
| import os | |
| import sqlite3 | |
| import sys | |
| import time | |
| from pathlib import Path | |
| import httpx | |
| from seed_suite import make_suite | |
| # 包里脚本在 scripts/,schemas/ data/ 在上一级;GitHub 仓库是平铺的,两种都认 | |
| ROOT = Path(__file__).resolve().parent | |
| if not (ROOT / "schemas").is_dir(): | |
| ROOT = ROOT.parent | |
| API = "https://api.deepseek.com/chat/completions" | |
| # 测试库:出训练题时碰都不能碰,不然测试集就泄漏了 | |
| TEST_DBS = {"library", "gym", "hard_retail", "transit"} | |
| PROMPT = """你在给一个中文 Text-to-SQLite 数据集出题。 | |
| 下面是数据库的建表语句: | |
| {ddl} | |
| 出 {n} 道中文问题,并为每题写一条 SQLite 查询。要求: | |
| - 问题要像真人问的,口语化,可以有「上个月」「最多的前五个」这种自然表达 | |
| - 覆盖不同难度:简单筛选、聚合、分组、多表连接、子查询、排序取前几 | |
| - 只能用 SELECT,一条语句,不要解释、不要分号结尾以外的多余内容 | |
| - 表名和列名必须严格使用上面出现过的 | |
| - 不要问需要 INSERT/UPDATE/DELETE 的问题 | |
| - 大约五分之一的题要牵涉 CTE(WITH)、窗口函数或者日期计算,别全是简单查询 | |
| - 日期字段是 TEXT,格式 'YYYY-MM-DD' 或 'YYYY-MM-DD HH:MM:SS',用 SQLite 的 date()/strftime() | |
| - 问题里不要出现表名、列名或 SQL 片段,全用业务说法;问「前几名」「最多的」要能有唯一答案 | |
| - 一对多连接后再 SUM 父表的金额会重复计算,这种要先在子查询/CTE 里聚合好再连 | |
| 只输出 JSON,格式: | |
| {{"items": [{{"question": "中文问题", "sql": "SELECT ..."}}]}}""" | |
| def api_key() -> str: | |
| """DeepSeek 的 key 从环境变量 DEEPSEEK_API_KEY 读,不写进代码和仓库。""" | |
| key = os.environ.get("DEEPSEEK_API_KEY", "").strip() | |
| if not key: | |
| raise SystemExit("先设置环境变量 DEEPSEEK_API_KEY") | |
| return key | |
| def post_with_retry(payload: dict, key: str, timeout: int = 180) -> str: | |
| """调 DeepSeek,返回回复正文。偶尔超时或 5xx,重试 3 次、一次比一次等得久;还不行就抛出去。""" | |
| for attempt in range(3): | |
| try: | |
| r = httpx.post(API, json=payload, timeout=timeout, headers={"Authorization": f"Bearer {key}"}) | |
| r.raise_for_status() | |
| return r.json()["choices"][0]["message"]["content"] | |
| except (httpx.HTTPError, KeyError, ValueError) as exc: | |
| if attempt == 2: | |
| raise | |
| print(f"请求失败({exc}),{10 * (attempt + 1)} 秒后重试", file=sys.stderr) | |
| time.sleep(10 * (attempt + 1)) | |
| raise AssertionError("unreachable") | |
| def ask(ddl: str, n: int, key: str) -> list[dict]: | |
| payload = { | |
| "model": "deepseek-chat", | |
| "messages": [ | |
| {"role": "system", "content": "你只输出 JSON。"}, | |
| {"role": "user", "content": PROMPT.format(ddl=ddl, n=n)}, | |
| ], | |
| "temperature": 1.0, | |
| "max_tokens": 4000, | |
| "response_format": {"type": "json_object"}, | |
| } | |
| try: | |
| return json.loads(post_with_retry(payload, key)).get("items") or [] | |
| except json.JSONDecodeError: | |
| print("这一轮 JSON 坏了,跳过", file=sys.stderr) | |
| return [] | |
| def check(sql: str, conn: sqlite3.Connection) -> tuple[bool, str]: | |
| s = sql.strip().rstrip(";") | |
| # WITH 开头的是 CTE,也要留;以前只认 select,老师写的 CTE 全被扔了,训练集里一条 CTE 都没有 | |
| if not s.lower().startswith(("select", "with")): | |
| return False, "不是 select" | |
| if ";" in s: | |
| return False, "多条语句" | |
| try: | |
| conn.execute(s).fetchall() | |
| return True, "" | |
| except Exception as exc: # noqa: BLE001 | |
| return False, str(exc) | |
| def main() -> None: | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--schema", required=True) | |
| ap.add_argument("--n", type=int, default=20) | |
| ap.add_argument("--rounds", type=int, default=1) | |
| ap.add_argument("--out", default="") | |
| args = ap.parse_args() | |
| schema_path = Path(args.schema) if Path(args.schema).exists() else ROOT / args.schema | |
| ddl = schema_path.read_text(encoding="utf-8") | |
| dbname = schema_path.stem | |
| if dbname in TEST_DBS: | |
| raise SystemExit(f"{dbname} 是测试库,不能拿来出训练题") | |
| conn = sqlite3.connect(":memory:") | |
| conn.executescript(ddl) | |
| conn.executescript(make_suite(dbname, [], 1)[0]) | |
| conn.execute("PRAGMA query_only = ON") # WITH ... DELETE 这种也别想改库 | |
| key = api_key() | |
| out_path = Path(args.out) if args.out else ROOT / "data" / "new" / f"{dbname}.jsonl" | |
| out_path.parent.mkdir(parents=True, exist_ok=True) | |
| kept = bad = 0 | |
| with out_path.open("a", encoding="utf-8") as f: | |
| for r in range(args.rounds): | |
| items = ask(ddl, args.n, key) | |
| for it in items: | |
| q = str(it.get("question") or "").strip() | |
| sql = str(it.get("sql") or "").strip() | |
| if not q or not sql: | |
| continue | |
| ok, err = check(sql, conn) | |
| if not ok: | |
| bad += 1 | |
| print(f"丢弃({err[:60]}):{sql[:80]}", file=sys.stderr) | |
| continue | |
| f.write(json.dumps({"db": dbname, "question": q, "sql": sql.rstrip(";")}, | |
| ensure_ascii=False) + "\n") | |
| kept += 1 | |
| print(f"第 {r+1} 轮:累计保留 {kept},丢弃 {bad}", flush=True) | |
| print(f"完成:{out_path},保留 {kept} 条,丢弃 {bad} 条") | |
| if __name__ == "__main__": | |
| main() | |