Files
MeltMind/docs/gre24-v5-stored-score-display-2026-11-16.md
T

40 lines
4.3 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Gre24 v5:已有评分的阈值及配对差展示补充
按Claude原串1794782386.838149要求,仅从固定 `scores.json` 读取既有逐件标量并汇总。原评分SHA `af2aa9fdb207c86c4ff91a75a33e84734e3995f37fd316ae394bfd231f57a674` 前后不变,原summary、预测、模型不改写;无新训练、预测、评分、bootstrap或标签读取。三分母min/median/max原表见已发布 `docs/gre24-v5-reproduction-results-2026-11-16.md`(9bf495a)。
## RRMSE严格小于0.1的比例
每格为通过件数/该split有效评分件数及比例;614成功/10准备超时的固定排除不变,五split分别有效307/308/306/305/308,缺5/4/6/7/4。分母是各split有效评分件数,不是312或节点数。五split比例等权平均,不先池化各split样本;无undefined评分。全部数值均来自原标量,严格小于0.1,不是小于等于。
| 分母 | 模型 | split0 | split1 | split2 | split3 | split4 | 五split比例平均 |
| --- | --- | --- | --- | --- | --- | --- | ---: |
| RMS(y) | naive | 174/307 (56.6775%) | 190/308 (61.6883%) | 173/306 (56.5359%) | 180/305 (59.0164%) | 181/308 (58.7662%) | 58.536882% |
| RMS(y) | GBRT | 238/307 (77.5244%) | 235/308 (76.2987%) | 226/306 (73.8562%) | 235/305 (77.0492%) | 232/308 (75.3247%) | 76.010639% |
| RMS(y) | FNN | 222/307 (72.3127%) | 220/308 (71.4286%) | 214/306 (69.9346%) | 207/305 (67.8689%) | 231/308 (75.0000%) | 71.308954% |
| max(y) | naive | 272/307 (88.5993%) | 272/308 (88.3117%) | 261/306 (85.2941%) | 267/305 (87.5410%) | 262/308 (85.0649%) | 86.962215% |
| max(y) | GBRT | 293/307 (95.4397%) | 291/308 (94.4805%) | 291/306 (95.0980%) | 292/305 (95.7377%) | 295/308 (95.7792%) | 95.307045% |
| max(y) | FNN | 288/307 (93.8111%) | 287/308 (93.1818%) | 285/306 (93.1373%) | 281/305 (92.1311%) | 286/308 (92.8571%) | 93.023688% |
| mean(y) | naive | 145/307 (47.2313%) | 158/308 (51.2987%) | 150/306 (49.0196%) | 157/305 (51.4754%) | 157/308 (50.9740%) | 49.999803% |
| mean(y) | GBRT | 207/307 (67.4267%) | 206/308 (66.8831%) | 203/306 (66.3399%) | 209/305 (68.5246%) | 210/308 (68.1818%) | 67.471221% |
| mean(y) | FNN | 193/307 (62.8664%) | 197/308 (63.9610%) | 187/306 (61.1111%) | 170/305 (55.7377%) | 200/308 (64.9351%) | 61.722274% |
## 同件配对的naive与GBRT差
每个split按相同case配对,先算 **Δ = RRMSE_naive − RRMSE_GBRT**,再对该split的Δ取中位数。正值表示GBRT误差较低;这不是两个模型各自中位数相减。五split配对数依次307/308/306/305/308,原失败排除不补替;逐对原数与Δ也在补充JSON内。数值为比例,非百分比。
| 分母 | split0 median(Δ) | split1 median(Δ) | split2 median(Δ) | split3 median(Δ) | split4 median(Δ) |
| --- | ---: | ---: | ---: | ---: | ---: |
| RMS(y) | 0.018514102312 | 0.017794192390 | 0.019783545436 | 0.017345627252 | 0.020289906179 |
| max(y) | 0.011086873306 | 0.010534568802 | 0.011531321491 | 0.010271705431 | 0.011972139328 |
| mean(y) | 0.021249401762 | 0.020732402248 | 0.023199691516 | 0.019992931523 | 0.023283182183 |
## 参考事实、实现选择与未能确认的内容
论文参考统计:Claude1794782386.838149报告原文300/314约95.5%,并说明原文没有定义RRMSE分母。本补充引用其团队原文核读记录,不声称本轮重新阅读原论文或确认其分母。
本次实现选择:明确列出三分母,RMS(y)为原固定主指标;仍为v5、t_set时间尺度、fill-only早停FNN及R1补充离散化。max(y)下GBRT阈值比例平均95.307045%,数值接近上述参考;主RMS(y)下为76.010639%。这说明同一批既有误差采用不同分母会产生很不同的展示,不能据接近某参考值反向选择主指标。
未能确认:原论文实际采用何种分母、v4/v5差异、t_set与V/Q、FNN多输出及离散化的独立影响均不能由这张表识别,10较大件超时排除与近重复跨split亦未消除。相近百分比不是忠实复现认证;本补充没有把总体差距归因给某一个候选因素。
新增展示汇总实际CPU **0.165208997秒**,GPU新增0。原完成账本136620.981045秒不改写;单列展示费用后累计 **136621.146253997秒/216000秒**。原GPU876.330852390秒/7200秒不变。费用回执 `stored-score-display-receipt-2026-11-16.json`,完整结果 `stored-score-display-supplement-2026-11-16.json` 均位于v5-models-continuation结果根。