RoyirinRosmontis/CareerNexus
1
1
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

真实数据集扩产与被动语料接入

Browse Source 
- scripts/expand_real_dataset.py：AutoWeakS 全量 2215 条 + 500 条分层评测集 + 简历扩量（seed=2026 可复现）
- pipeline 被动语料机制（passive_sources）：真实无日期 JD 参与接入/清洗/抽取验证（2915 条命中率 99.8%），不进图谱/演化/发现
- 新岗位发现第四指标'差异度'与双年窗口增长率口径
- 评测新增 real_jd_500 维度（99.91%）；测试会话隔离 jd_extra_dirs 防止本机配置拖慢套件
This commit is contained in:
RoyirinRosmontis
2026-08-03 15:07:54 +08:00
parent 441d555caf
commit 1722c00249
9 changed files with 412 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/evaluate.py
+32 -7
View File
@@ -9,8 +9,9 @@
两个真实数据维度（data/real/，与合成集分开报告、不混算）：
4. 真实 JD（AutoWeakS 初标+抽检 silver gold）—— 对 jd_sample.jsonl 重新抽取，
与 gold_answers.json 算技能/场景字段级 F1；
5. 真实简历（AIStudio CC0）—— 对 50 份简历解析，姓名/电话命中率 + 技能 F1。
与 gold_answers.json 算技能/场景字段级 F1；另按 jd_eval_500.jsonl /
gold_answers_500.json 报告 500 条分层口径（real_jd_500，包含原 200 条）；
5. 真实简历（AIStudio CC0）—— 对 200 份简历解析，姓名/电话命中率 + 技能 F1。
输出 reports/metrics.json 与控制台报告。
"""
@@ -115,15 +116,14 @@ def eval_match() -> dict:
return {"accuracy": round(hit / len(pairs), 4), "pairs": len(pairs), "detail": detail}
def eval_real_jd() -> dict:
def _eval_real_jd(records_path: Path, gold_path: Path) -> dict:
"""真实 JD（AutoWeakS）：从文本重新抽取，与 silver gold（初标+抽检）字段级比对。
真实 JD 是自由文本、无编号版面，gold 的 responsibilities/preferred_skills 为空，
口径 = 全文词表技能（required_skills）+ 应用场景 + 启发式岗位名，见 data/real/README.md。
"""
records, errors = load_jd_file(ROOT / "data" / "real" / "autoweaks_jd" / "jd_sample.jsonl")
gold = json.loads(
(ROOT / "data" / "real" / "autoweaks_jd" / "gold_answers.json").read_text(encoding="utf-8"))
records, errors = load_jd_file(records_path)
gold = json.loads(gold_path.read_text(encoding="utf-8"))
extractor = OfflineExtractor(load_lexicon())
fields = ["required_skills", "preferred_skills", "scenarios"]
@@ -143,7 +143,9 @@ def eval_real_jd() -> dict:
detail, active_f1 = {}, []
for f in fields:
if not gold_n[f] and not pred[f]:
# 真实 JD 的 gold 口径：responsibilities/preferred_skills 为空即不参与评分
# （即使抽取器从"优先考虑"字样产出加分候选，也不计入，口径见 data/real/README.md）
if not gold_n[f]:
detail[f] = None # 该字段在真实语料中无标注口径，不参与
continue
p, r, f1 = prf(tp[f], pred[f], gold_n[f])
@@ -157,6 +159,20 @@ def eval_real_jd() -> dict:
"fields": detail, "samples": len(records), "bad_lines": len(errors)}
def eval_real_jd() -> dict:
"""真实 JD 200 条口径（历史基线）。"""
return _eval_real_jd(
ROOT / "data" / "real" / "autoweaks_jd" / "jd_sample.jsonl",
ROOT / "data" / "real" / "autoweaks_jd" / "gold_answers.json")
def eval_real_jd_500() -> dict:
"""真实 JD 500 条分层口径（含原 200 条，见 scripts/expand_real_dataset.py）。"""
return _eval_real_jd(
ROOT / "data" / "real" / "autoweaks_jd" / "jd_eval_500.jsonl",
ROOT / "data" / "real" / "autoweaks_jd" / "gold_answers_500.json")
def eval_real_resume() -> dict:
"""真实简历（AIStudio CC0）：姓名/电话命中率（脱敏形态比对）+ 技能 F1。"""
gold = json.loads(
@@ -194,6 +210,7 @@ def main() -> None:
resume = eval_resume()
match = eval_match()
real_jd = eval_real_jd()
real_jd_500 = eval_real_jd_500()
real_resume = eval_real_resume()
metrics = {
"jd_extraction_accuracy": jd["accuracy"],
@@ -203,6 +220,7 @@ def main() -> None:
"resume_detail": resume["detail"],
"match_detail": match["detail"],
"real_jd": real_jd,
"real_jd_500": real_jd_500,
"real_resume": real_resume,
}
out = ROOT / "reports"
@@ -225,6 +243,13 @@ def main() -> None:
print("真实数据维度（data/real/，与合成集分开报告）")
print(f"真实 JD（AutoWeakS 初标+抽检）: 综合 F1 {real_jd['accuracy']:.2%}"
f" (样本 {real_jd['samples']} 条)")
print(f"真实 JD（500 条分层） : 综合 F1 {real_jd_500['accuracy']:.2%}"
f" (样本 {real_jd_500['samples']} 条)")
for f, d in real_jd_500["fields"].items():
if d is None:
print(f" - {f:<18} (真实语料无标注口径，不参与)")
else:
print(f" - {f:<18} P={d['precision']:.2%} R={d['recall']:.2%} F1={d['f1']:.2%}")
for f, d in real_jd["fields"].items():
if d is None:
print(f" - {f:<18} (真实语料无标注口径，不参与)")