fix(metrics): t_c70f385c 指标口径修复 — recall_hit_rate 退役 + gap_closure_rate 断言 + 最小样本门 N=20
- recall_hit_rate 退役:分子原为 core.go 的 len(results)>0(恒=1、无区分度), 「返回且被使用」语义已由 recall_usefulness_rate 承担(207 样本 0.98)。 Metrics()/distributed/metrics-gauge/executor/tuning/CLI/daily-check 全链移除该键。 - gap_closure_rate 统一为「缺口个数」口径:Close() 加 !g.Closed 守卫(与 RecordHit 一致), 根治同一缺口重复关闭导致 ClosedGaps 虚增、越界 4.5;Metrics() 加越界断言 gapRate>1 → -1。 - 所有率类指标加最小样本门 minRateSamples=20:样本<20 ⇒ -1 + valid=0, 杜绝 auto_resolve_rate 2 样本判满分。 - 测试:metrics_retire_regression_test.go 用生产同款计数复现改前/改后 JSON; selfoptimize 全套通过。
This commit is contained in:
@@ -26,13 +26,15 @@ METRICS=$(curl -s $BASE/api/v1/metrics/self -H "X-API-Key: $API_KEY" 2>/dev/null
|
||||
echo "$METRICS" | python3 -c "
|
||||
import sys,json
|
||||
d=json.load(sys.stdin)
|
||||
hit = d.get('recall_hit_rate',0)
|
||||
loss = d.get('avg_distill_loss',0)
|
||||
gap = d.get('gap_closure_rate',0)
|
||||
auto = d.get('auto_resolve_rate',0)
|
||||
print(f'命中率: {hit:.0%}, 蒸馏损失: {loss:.2f}, gap闭合: {gap}/天, auto_resolve: {auto}')
|
||||
# 2026-10-17:recall_hit_rate 已退役(恒=1、无区分度)。召回质量改用 recall_usefulness_rate
|
||||
#(返回且被使用);-1 表示样本不足/无负反馈 = 数据不可解读,跳过告警,不误报。
|
||||
hit = d.get('recall_usefulness_rate', -1)
|
||||
loss = d.get('avg_distill_loss', 0)
|
||||
gap = d.get('gap_closure_rate', 0)
|
||||
auto = d.get('auto_resolve_rate', 0)
|
||||
print(f'召回有用率: {hit:.0%}, 蒸馏损失: {loss:.2f}, gap闭合: {gap}/天, auto_resolve: {auto}')
|
||||
if loss > 0.3: print('⚠️ avg_distill_loss 超标')
|
||||
if hit < 0.5: print('⚠️ recall_hit_rate 低于阈值')
|
||||
if hit >= 0 and hit < 0.5: print('⚠️ recall_usefulness_rate 低于阈值')
|
||||
" 2>/dev/null
|
||||
|
||||
# 4. 蒸馏状态
|
||||
|
||||
Reference in New Issue
Block a user