# -*- coding: utf-8 -*- """資料集品質驗證。 檢查項目： 1. 欄位與型別是否對齊 twinkle-ai/tw-function-call-reasoning-10k 2. answer 是否為合法 JSON，且呼叫的函式都在 tools 清單內 3. 參數名稱是否都存在於該函式的 schema 4. messages 三段是否與 query_zhtw / think / answer 完全一致（可還原） 5. 參數值是否「有據可循」——能在題目文字裡找到來源，而非模型臆測 6. train / test 是否有完全重複的樣本 """ import json import re import sys from collections import Counter SCHEMA = ["id", "query", "tools", "query_zhtw", "think", "answer", "messages"] TYPES = {"id": int, "query": str, "tools": str, "query_zhtw": str, "think": str, "answer": str, "messages": list} # enum 參數在題目中常以中文表達，驗證溯源時需一併認列 ENUM_ZH = { "standard": "標準扣除額", "itemized": "列舉扣除額", "single": "單身", "married_joint": "夫妻合併申報", "fixed": "固定利率", "floating": "機動利率", "simple": "單利存本取息", "monthly_compound": "整存整付按月複利", "equal_total": "本息平均攤還", "equal_principal": "本金平均攤還", "income_replacement": "所得替代法", "needs_based": "需求分析法", "stable": "穩定", "moderate": "普通", "volatile": "不穩定", "mortgage_application": "房貸申請", "credit_card_application": "信用卡申請", "self_inquiry": "本人查詢", "personal_loan": "信用貸款", "buy": "買進", "sell": "賣出", "stock": "股票", "etf": "股票型ETF", "warrant": "權證", "corporate_bond_etf": "債券型ETF", "dividend": "股利", "bonus": "年終獎金", "rent": "租金", "professional_fee": "執行業務所得", "part_time_salary": "兼職薪資", "interest": "利息", "cash_buy": "現金買入", "cash_sell": "現金賣出", "spot_buy": "即期買入", "spot_sell": "即期賣出", "rediscount": "重貼現率", "accommodations_with_collateral": "擔保放款融通利率", "accommodations_without_collateral": "短期融通利率", "monthly": "月配", "quarterly": "季配", "semiannual": "半年配", "cash_deposit": "現金存入", "cash_withdrawal": "現金提領", "wire_transfer": "電匯", "fx_exchange": "結匯", "individual": "自然人", "corporate": "法人", "trust": "信託", "beginner": "完全新手", "intermediate": "有基礎", "professional": "從業人員", "TAIEX": "台股加權指數", "TPEx": "櫃買指數", "TWSE50": "台灣50指數", "TAIEX_TR": "發行量加權股價報酬指數", "FITX": "台指期近月", "USD": "美元", "JPY": "日圓", "EUR": "歐元", "CNY": "人民幣", "AUD": "澳幣", "HKD": "港幣", "GBP": "英鎊", "KRW": "韓元", "zh-TW": "繁體中文", "close": "收盤", "change_pct": "漲跌", "open": "開", "high": "高", "low": "低", "volume": "量", } def load(p): with open(p, encoding="utf-8") as f: return [json.loads(l) for l in f if l.strip()] def grounded(val, text): """參數值是否能在題目文字中找到來源（容許萬元、百分比等常見換算）。""" if isinstance(val, bool): return True # 布林多由中文語意表達（如「當沖」「自住」），另由模板保證 if isinstance(val, (int, float)) and not isinstance(val, bool): s = ("%g" % val) cands = {s, str(val), "%g" % abs(val)} if isinstance(val, int): if val % 10000 == 0: cands.add(str(val // 10000)) cands.add("{:,}".format(val)) else: cands.add("%g" % (val * 100)) return any(c in text for c in cands) if isinstance(val, str): return val in text or ENUM_ZH.get(val, "\x00") in text if isinstance(val, list): return all(grounded(v, text) for v in val) return True def check(rows, name, errors, stats): for r in rows: rid = r.get("id", "?") if list(r.keys()) != SCHEMA: errors.append("%s#%s 欄位不符：%s" % (name, rid, list(r.keys()))) continue for k, t in TYPES.items(): if not isinstance(r[k], t): errors.append("%s#%s 欄位 %s 型別應為 %s" % (name, rid, k, t.__name__)) try: tools = json.loads(r["tools"]) calls = json.loads(r["answer"]) except Exception as e: errors.append("%s#%s JSON 解析失敗：%s" % (name, rid, e)) continue tmap = {t["name"]: t for t in tools} for c in calls: if c["name"] not in tmap: errors.append("%s#%s 呼叫了不在 tools 清單的函式 %s" % (name, rid, c["name"])) continue params = tmap[c["name"]]["parameters"] for k, v in c["arguments"].items(): if k not in params: errors.append("%s#%s 參數 %s 不在 %s 的 schema" % (name, rid, k, c["name"])) stats["args_total"] += 1 if grounded(v, r["query_zhtw"] + " " + r["query"]): stats["args_grounded"] += 1 else: stats["ungrounded"].append((rid, c["name"], k, v)) m = r["messages"] if len(m) != 3 or [x["role"] for x in m] != ["system", "user", "assistant"]: errors.append("%s#%s messages 結構不是 system/user/assistant" % (name, rid)) continue if m[1]["content"] != r["query_zhtw"]: errors.append("%s#%s messages[1] 與 query_zhtw 不一致" % (name, rid)) if r["think"] not in m[2]["content"]: errors.append("%s#%s think 未出現在 assistant 訊息" % (name, rid)) found = re.findall(r"\n(.*?)\n", m[2]["content"], re.S) if len(found) != len(calls): errors.append("%s#%s tool_call 數量(%d)與 answer(%d)不符" % (name, rid, len(found), len(calls))) else: for got, want in zip(found, calls): if json.loads(got) != want: errors.append("%s#%s tool_call 內容與 answer 不一致" % (name, rid)) for t in tools: if ('"%s"' % t["name"]) not in m[0]["content"]: errors.append("%s#%s system 訊息缺少工具 %s" % (name, rid, t["name"])) stats["calls"] += len(calls) stats["tools_offered"] += len(tools) def main(): tr = load(sys.argv[1] if len(sys.argv) > 1 else "data/train.jsonl") te = load(sys.argv[2] if len(sys.argv) > 2 else "data/test.jsonl") errors = [] stats = Counter() stats["ungrounded"] = [] check(tr, "train", errors, stats) check(te, "test", errors, stats) ids = [r["id"] for r in tr + te] if len(set(ids)) != len(ids): errors.append("id 有重複") tr_keys = {(r["query_zhtw"], r["answer"]) for r in tr} dup = [r["id"] for r in te if (r["query_zhtw"], r["answer"]) in tr_keys] if dup: errors.append("test 有 %d 筆與 train 完全重複：%s" % (len(dup), dup[:10])) n = len(tr) + len(te) print("=" * 66) print("資料集驗證報告") print("=" * 66) print("train / test : %d / %d (%.1f%% : %.1f%%)" % (len(tr), len(te), 100 * len(tr) / n, 100 * len(te) / n)) print("函式呼叫總數 : %d（平均每題 %.2f 次）" % (stats["calls"], stats["calls"] / n)) print("平均提供工具數 : %.2f" % (stats["tools_offered"] / n)) print("參數溯源率 : %d/%d = %.2f%%" % (stats["args_grounded"], stats["args_total"], 100 * stats["args_grounded"] / max(1, stats["args_total"]))) print("train/test 完全重複 : %d 筆" % len(dup)) print("-" * 66) if stats["ungrounded"]: print("題目中找不到來源的參數（前 10 筆）：") for u in stats["ungrounded"][:10]: print(" #%s %s.%s = %r" % u) print() if errors: print("發現 %d 項錯誤：" % len(errors)) for e in errors[:25]: print(" x", e) sys.exit(1) print("結果：通過，無結構性錯誤。") if __name__ == "__main__": main()