| """用老师再检查一遍：这条 SQL 真的回答了那个问题吗？ |
| |
| 执行能通过不等于答对，比如问「每个年级的平均工资」却给个 join 出来的 salary 列表。 |
| 这里让模型逐条判断，判错的直接剔掉。 |
| |
| python verify.py --data data/school.jsonl |
| """ |
|
|
| import argparse |
| import json |
| import os |
| import re |
| import sqlite3 |
| from pathlib import Path |
|
|
| from teacher import post_with_retry |
|
|
| |
| ROOT = Path(__file__).resolve().parent |
| if not (ROOT / "schemas").is_dir(): |
| ROOT = ROOT.parent |
| API = "https://api.deepseek.com/chat/completions" |
|
|
| def find_schema(db: str) -> Path: |
| for p in (ROOT / "schemas").rglob(f"{db}.sql"): |
| return p |
| raise FileNotFoundError(db) |
|
|
|
|
| PROMPT = """下面是一组「中文问题 + SQLite 查询」。逐条判断这条 SQL 是否准确回答了问题。 |
| |
| {rows} |
| |
| 判断标准： |
| - 该查的字段、该做的聚合、该有的过滤条件都在，才算对 |
| - 多查了无关字段但仍是同一件事，算对 |
| - 少条件、聚合错、表连错、语义跑偏，算错 |
| - 问题有歧义时，按最自然的理解判断 |
| |
| 只输出 JSON：{{"verdicts": [{{"i": 1, "ok": true, "why": "一句话理由"}}]}}""" |
|
|
|
|
| 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 judge(batch: list[dict], ddl: str, key: str) -> dict[int, bool]: |
| rows = "\n".join( |
| f'{i+1}. 问题：{b["question"]}\n SQL：{b["sql"]}' for i, b in enumerate(batch) |
| ) |
| payload = { |
| "model": "deepseek-chat", |
| "messages": [ |
| {"role": "system", "content": "你只输出 JSON。"}, |
| {"role": "user", "content": f"建表语句：\n{ddl}\n\n{PROMPT.format(rows=rows)}"}, |
| ], |
| "temperature": 0.0, |
| "max_tokens": 2000, |
| "response_format": {"type": "json_object"}, |
| } |
| try: |
| data = json.loads(post_with_retry(payload, key)) |
| except json.JSONDecodeError: |
| |
| print(f"这一批 {len(batch)} 条的判定 JSON 坏了，先保留", flush=True) |
| return {} |
| out = {} |
| for v in data.get("verdicts") or []: |
| try: |
| out[int(v["i"])] = bool(v.get("ok")) |
| except Exception: |
| continue |
| return out |
|
|
|
|
| def main() -> None: |
| ap = argparse.ArgumentParser() |
| ap.add_argument("--data", required=True) |
| ap.add_argument("--batch", type=int, default=15) |
| ap.add_argument("--keep-all", action="store_true") |
| args = ap.parse_args() |
|
|
| path = Path(args.data) |
| items = [json.loads(l) for l in path.read_text(encoding="utf-8").splitlines() if l.strip()] |
| key = api_key() |
| kept = [] |
| for start in range(0, len(items), args.batch): |
| batch = items[start:start + args.batch] |
| ddl = find_schema(batch[0]["db"]).read_text(encoding="utf-8") |
| verdict = judge(batch, ddl, key) |
| for i, it in enumerate(batch, 1): |
| if verdict.get(i, True): |
| kept.append(it) |
| else: |
| print(f"剔除：{it['question']} | {it['sql'][:70]}", flush=True) |
| print(f"{min(start+args.batch, len(items))}/{len(items)} 留下 {len(kept)}", flush=True) |
|
|
| path.write_text("\n".join(json.dumps(k, ensure_ascii=False) for k in kept) + "\n", encoding="utf-8") |
| print(f"完成：{len(items)} → {len(kept)}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
