Add v4 stored-score threshold fractions and paired medians

This commit is contained in:
Frank14f committed 2026-11-16 23:42:20 +08:00
1 parent 271c47a698
commit 7bf1f922e3
5 files changed
+531

No files matched your search

@@ -10,6 +10,8 @@
## 已有评分文件的数值
既有分数的三分母严格 `<0.1` 比例、逐split同件配对 naive−GBRT 中位数,已另列于[展示补充](gre24-v4-stored-score-display-2026-11-16.md)。仅汇总原标量,未重新评分;原14个准备失败及所有解释边界保留。
下表直接摘录各split先取中位数、再对五个split取平均的原摘要,未重新聚合预测或评分。原域为全部保留tet节点(含壁面),不是S2查询;这些值不能直接用于R4/S2组件分组归因。
|分母|naive|GBRT|FNN|
@@ -0,0 +1,37 @@
# Gre24 v4:已有评分的阈值及配对差展示补充
按Claude1794839041.521639要求,仅汇总固定 scores.json 内已有逐件标量。原评分SHA256为 `7308e1b74d96e1f003766b72fc751c2cb0eedb9956d906d2446924015fcd04a9`,读取前后相同;未读取标签、预测数组或模型,未重新拟合、预测、评分或bootstrap。原评分和终态账不改。
## RRMSE严格小于0.1的比例
五次独立随机315测试/314训练划分,不是五折交叉验证。有效测试依次304/311/308/309/305,缺失11/4/7/6/10;1575次测试访问含38个准备失败行。比例分母为各split有效评分件数,五split比例等权平均,不按节点加权或先池化;14个准备失败及原排除保持不变。
| 分母 | 模型 | split0 | split1 | split2 | split3 | split4 | 五split比例平均 |
| --- | --- | --- | --- | --- | --- | --- | ---: |
| RMS(y) | naive | 178/304 (58.5526%) | 185/311 (59.4855%) | 191/308 (62.0130%) | 185/309 (59.8706%) | 186/305 (60.9836%) | 60.181061% |
| RMS(y) | GBRT | 231/304 (75.9868%) | 239/311 (76.8489%) | 246/308 (79.8701%) | 246/309 (79.6117%) | 243/305 (79.6721%) | 78.397926% |
| RMS(y) | FNN | 222/304 (73.0263%) | 226/311 (72.6688%) | 226/308 (73.3766%) | 220/309 (71.1974%) | 227/305 (74.4262%) | 72.939078% |
| max(y) | naive | 259/304 (85.1974%) | 275/311 (88.4244%) | 268/308 (87.0130%) | 272/309 (88.0259%) | 265/305 (86.8852%) | 87.109186% |
| max(y) | GBRT | 285/304 (93.7500%) | 295/311 (94.8553%) | 292/308 (94.8052%) | 294/309 (95.1456%) | 292/305 (95.7377%) | 94.858767% |
| max(y) | FNN | 284/304 (93.4211%) | 292/311 (93.8907%) | 287/308 (93.1818%) | 291/309 (94.1748%) | 289/305 (94.7541%) | 93.884480% |
| mean(y) | naive | 155/304 (50.9868%) | 161/311 (51.7685%) | 164/308 (53.2468%) | 155/309 (50.1618%) | 159/305 (52.1311%) | 51.659009% |
| mean(y) | GBRT | 204/304 (67.1053%) | 216/311 (69.4534%) | 217/308 (70.4545%) | 215/309 (69.5793%) | 215/305 (70.4918%) | 69.416855% |
| mean(y) | FNN | 189/304 (62.1711%) | 195/311 (62.7010%) | 189/308 (61.3636%) | 193/309 (62.4595%) | 204/305 (66.8852%) | 63.116089% |
## 同件配对的naive与GBRT差
先在每件上计算 Δ = RRMSE_naive − RRMSE_GBRT,再在各split取Δ中位数;不是两模型中位数相减。正值表示GBRT误差较低。配对数量为304/311/308/309/305,数值为比例而非百分比。
| 分母 | split0 median(Δ) | split1 median(Δ) | split2 median(Δ) | split3 median(Δ) | split4 median(Δ) |
| --- | ---: | ---: | ---: | ---: | ---: |
| RMS(y) | 0.019882618405 | 0.018171540025 | 0.018917791542 | 0.019676381359 | 0.017812751847 |
| max(y) | 0.011888508286 | 0.010736216578 | 0.011273753025 | 0.011581302653 | 0.011069599973 |
| mean(y) | 0.023021678969 | 0.021097379048 | 0.021652869330 | 0.022773028567 | 0.020617144249 |
## 解释边界与费用
RMS(y)仍为内部预登记主指标;max(y)是保留tet域最后到达时间,作为论文total fill time的代理,作者评分代码尚未取得。三分母完整保留,不据数值接近改主指标。失败覆盖偏差方向未知,v4/v5摘要相近不能建立版本因果配对;t_set替代缺失Q、R1离散补充、fill-only/早停FNN和随机几何近重复限制均保留。
这是既有分数的描述性展示,不新增统计检验、训练或版本因果结论。原始结果见[固定结果报告](gre24-v4-reproduction-results-2026-11-16.md),独立审计见[MeltAssist审计](meltassist-v4-terminal-audit-2026-11-16.md)。完整计数与配对中位数见[补充JSON](../artifacts/gre24-v4-stored-score-display-2026-11-16/supplement.json)。
此次展示汇总CPU 0.299509830秒单列;原科学账140042.739554秒不改,加本次展示为140043.039063830/216000秒。先前独立审计费用仍另列,非这里的全活动总账。GPU新增0。费用与源码哈希见[回执](../artifacts/gre24-v4-stored-score-display-2026-11-16/receipt.json)。