Datasets:
Download scripts/ask.py from YukiKaze1337/cn-sqlite-sft: direct link, hf CLI and curl.
- Browser
- Download file 4.91 kB
-
https://huggingface.co/datasets/YukiKaze1337/cn-sqlite-sft/resolve/main/scripts/ask.py
- Command line
-
hf download hf://datasets/YukiKaze1337/cn-sqlite-sft/scripts/ask.py
-
curl -L -o ask.py https://huggingface.co/datasets/YukiKaze1337/cn-sqlite-sft/resolve/main/scripts/ask.py
4.91 kB
| """给一个 SQLite 库和一句中文,返回 SQL 和查询结果。 | |
| python scripts/ask.py --db shop.sqlite --question "上个月线上渠道取消了多少单?" | |
| python scripts/ask.py --db shop.sqlite # 不给问题就一行一行地问,空行退出 | |
| 模型默认用 out/merged;只有 LoRA 的话:--model Qwen/Qwen3-4B-Instruct-2507 --adapter out/lora。 | |
| 库以只读方式打开(mode=ro + query_only),模型写出 DELETE / UPDATE 也改不了数据。 | |
| 默认 SQL 跑不通就把报错发回去改,最多 2 次;有 GPU 时再一次写 5 条按执行结果投票,没有 GPU 就只写 1 条。 | |
| 第二轮模型在零售难题集上:直接生成 74%,出错重写 77%,投票加重写 82%(results/infer/summary.md)。 | |
| 建表语句和「部分字段的取值」都从库里现读,和训练时的 prompt 格式一样(prompt.py)。 | |
| 把握不够就不给结果,只说明原因(infer.confidence):投票时 5 条里少于 3 条结果一致,或者改完还跑不通。 | |
| 在两个难题集上,这样会少答约两成的题,给出去的错答案从 41 题降到 17 题。--show-unsure 照样显示,标上「仅供参考」。 | |
| """ | |
| import argparse | |
| import sqlite3 | |
| import time | |
| import eval as ev | |
| from infer import Generator, answer, confidence | |
| from prompt import build, values_from_db | |
| def open_readonly(path: str) -> sqlite3.Connection: | |
| conn = sqlite3.connect(f"file:{path}?mode=ro", uri=True) | |
| conn.execute("PRAGMA query_only = ON") | |
| return conn | |
| def schema_of(conn: sqlite3.Connection) -> str: | |
| return "\n\n".join(sql + ";" for (sql,) in conn.execute( | |
| "select sql from sqlite_master where type='table' and name not like 'sqlite_%' and sql is not null")) | |
| def show(conn: sqlite3.Connection, sql: str, limit: int) -> None: | |
| deadline = time.monotonic() + ev.QUERY_TIMEOUT | |
| conn.set_progress_handler(lambda: time.monotonic() > deadline, 10_000) | |
| try: | |
| cur = conn.execute(sql) | |
| rows = cur.fetchmany(limit + 1) | |
| except sqlite3.Error as exc: | |
| print("执行出错:", exc) | |
| return | |
| finally: | |
| conn.set_progress_handler(None, 0) | |
| print(" | ".join(d[0] for d in cur.description)) | |
| for r in rows[:limit]: | |
| print(" | ".join("NULL" if v is None else str(v) for v in r)) | |
| print(f"(只显示前 {limit} 行)" if len(rows) > limit else f"共 {len(rows)} 行") | |
| def main() -> None: | |
| ap = argparse.ArgumentParser() | |
| ap.add_argument("--db", required=True, help="SQLite 库文件") | |
| ap.add_argument("--question", default="") | |
| ap.add_argument("--model", default="out/merged") | |
| ap.add_argument("--adapter", default="") | |
| ap.add_argument("--repair", type=int, default=2) | |
| ap.add_argument("--vote", type=int, default=0, help="写几条投票;默认有 GPU 时 5 条,没有时 1 条(CPU 上写 5 条太慢)") | |
| ap.add_argument("--no-values", action="store_true", help="prompt 里不附「部分字段的取值」") | |
| ap.add_argument("--timeout", type=float, default=30.0, help="单条 SQL 最多跑几秒,大库调大") | |
| ap.add_argument("--rows", type=int, default=20, help="最多显示几行结果") | |
| ap.add_argument("--min-votes", type=int, default=3, help="投票时至少几条结果一致才给答案") | |
| ap.add_argument("--show-unsure", action="store_true", help="没把握时也显示 SQL 和结果,标上「仅供参考」") | |
| args = ap.parse_args() | |
| ev.QUERY_TIMEOUT = args.timeout | |
| conn = open_readonly(args.db) | |
| ddl = schema_of(conn) | |
| values = None if args.no_values else values_from_db(conn) | |
| gen = Generator(args.model, adapter=args.adapter) | |
| if args.vote <= 0: | |
| args.vote = 5 if gen.torch.cuda.is_available() else 1 | |
| n_tokens = len(gen.tok.apply_chat_template(build(ddl, "?", values), tokenize=True)) | |
| if n_tokens > 3000: | |
| print(f"注意:建表语句加取值有 {n_tokens} token,训练时最长的样本 2678 token,太长的库效果会变差") | |
| questions = [args.question] if args.question else iter(lambda: input("\n问题:").strip(), "") | |
| for q in questions: | |
| out = answer(gen, [build(ddl, q, values)], [conn], repair=args.repair, vote=args.vote)[0] | |
| report(conn, out, args.vote, args.min_votes, args.rows, args.show_unsure) | |
| def report(conn: sqlite3.Connection, out: dict, vote: int, min_votes: int, limit: int, show_unsure: bool) -> None: | |
| level, note = confidence(out, conn, vote, min_votes) | |
| unsure = level in ("low", "failed") | |
| print("\n" + note) | |
| if unsure and not show_unsure: | |
| return | |
| print(("\nSQL(仅供参考):" if unsure else "\nSQL:") + (out["sql"] or "(模型没写出 SQL)")) | |
| if out["repairs"]: | |
| print(f"(第一次写的跑不通,改了 {out['repairs']} 次)") | |
| if out["sql"] and level != "failed": | |
| show(conn, out["sql"], limit) | |
| if __name__ == "__main__": | |
| main() | |