Files
natural-memory-nm21/router_verdict_prod.md
T
WpyQwq 643e22ecb9 Natural Memory NM2.1: 记忆路由器分叉、数据集缺陷修复与全轴评测证据
- 引入 MemoryRouterXL 与 v5/v6 流式多线程训练/编码管线
- 修复 prepare_memory_router_dataset 候选池重建缺陷(mega 家族 3568x 加速,输出逐字节相同)
- 修复 v5 被破坏的拒答与多跳标签(train 未知样本 319 -> 16319,multi_hop 平均正例 1.00 -> 2.00)
- 同存储预算下 V2-128 v6 逐轴 22/22 通过:Top-1 41.12% -> 94.62%,未知拒答 0.00% -> 100.00%
- 记录三条被实测推翻的显然优化(logits_to_keep=1 反而慢 55%、XL 容量未带来收益)
- 记忆手术跨架构可移植性 14/14,读写关闭时与原生模型逐位相同
2026-09-19 11:11:31 +08:00

8.1 KiB

指标 最强旧基线 基线值 候选值 差值 判定
Top-1 正确率 V2-128 deployed(v3) 41.12% 94.14% +53.01pp 通过
Top-1 正确率 V2-128 deployed(v3) 41.12% 93.78% +52.66pp 通过
Top-1 正确率 V2-128 deployed(v3) 41.12% 93.81% +52.68pp 通过
Top-1 正确率 V2-128 deployed(v3) 41.12% 94.62% +53.49pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 88.58% +51.55pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 88.34% +51.32pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 88.38% +51.35pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 89.19% +52.16pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 96.48% +49.75pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 96.59% +49.86pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 96.78% +50.04pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 97.81% +51.07pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 97.15% +48.23pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 97.33% +48.41pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 97.78% +48.86pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 98.40% +49.49pp 通过
MRR V2-128 deployed(v3) 47.69% 95.77% +48.08pp 通过
MRR V2-128 deployed(v3) 47.69% 95.59% +47.90pp 通过
MRR V2-128 deployed(v3) 47.69% 95.70% +48.01pp 通过
MRR V2-128 deployed(v3) 47.69% 96.49% +48.80pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 95.37% +51.15pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 95.39% +51.17pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 95.53% +51.31pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 96.49% +52.27pp 通过
多跳证据全中(Top-3) V2-128 deployed(v3) 43.43% 96.22% +52.80pp 通过
多跳证据全中(Top-3) V2-128 deployed(v3) 43.43% 96.50% +53.07pp 通过
多跳证据全中(Top-3) V2-128 deployed(v3) 43.43% 96.70% +53.28pp 通过
多跳证据全中(Top-3) V2-128 deployed(v3) 43.43% 97.75% +54.32pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 95.45% +58.30pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 98.35% +61.20pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 98.75% +61.60pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 99.00% +61.85pp 通过
hop 正确率 V2-128 deployed(v3) 73.23% 100.00% +26.77pp 通过
hop 正确率 V2-128 deployed(v3) 73.23% 100.00% +26.77pp 通过
hop 正确率 V2-128 deployed(v3) 73.23% 100.00% +26.77pp 通过
hop 正确率 V2-128 deployed(v3) 73.23% 100.00% +26.77pp 通过
hop 欠预测率 V2-128 deployed(v3) 4.68% 0.00% -4.68pp 通过
hop 欠预测率 V2-128 deployed(v3) 4.68% 0.00% -4.68pp 通过
hop 欠预测率 V2-128 deployed(v3) 4.68% 0.00% -4.68pp 通过
hop 欠预测率 V2-128 deployed(v3) 4.68% 0.00% -4.68pp 通过
need F1 (门槛 0.50) V2-128 deployed(v3) 89.58% 100.00% +10.42pp 通过
need F1 (门槛 0.50) V2-128 deployed(v3) 89.58% 100.00% +10.42pp 通过
need F1 (门槛 0.50) V2-128 deployed(v3) 89.58% 100.00% +10.42pp 通过
need F1 (门槛 0.50) V2-128 deployed(v3) 89.58% 100.00% +10.42pp 通过
need 召回 (门槛 0.50) V2-128 deployed(v3) 99.82% 100.00% +0.18pp 通过
need 召回 (门槛 0.50) V2-128 deployed(v3) 99.82% 100.00% +0.18pp 通过
need 召回 (门槛 0.50) V2-128 deployed(v3) 99.82% 100.00% +0.18pp 通过
need 召回 (门槛 0.50) V2-128 deployed(v3) 99.82% 100.00% +0.18pp 通过
未知拒答率 (门槛 0.50) V2-128 deployed(v3) 0.00% 100.00% +100.00pp 通过
未知拒答率 (门槛 0.50) V2-128 deployed(v3) 0.00% 100.00% +100.00pp 通过
未知拒答率 (门槛 0.50) V2-128 deployed(v3) 0.00% 100.00% +100.00pp 通过
未知拒答率 (门槛 0.50) V2-128 deployed(v3) 0.00% 100.00% +100.00pp 通过
已知问题被误拒率 V2-128 deployed(v3) 0.18% 0.00% -0.18pp 通过
已知问题被误拒率 V2-128 deployed(v3) 0.18% 0.00% -0.18pp 通过
已知问题被误拒率 V2-128 deployed(v3) 0.18% 0.00% -0.18pp 通过
已知问题被误拒率 V2-128 deployed(v3) 0.18% 0.00% -0.18pp 通过
未知问题被误读率 V2-128 deployed(v3) 100.00% 0.00% -100.00pp 通过
未知问题被误读率 V2-128 deployed(v3) 100.00% 0.00% -100.00pp 通过
未知问题被误读率 V2-128 deployed(v3) 100.00% 0.00% -100.00pp 通过
未知问题被误读率 V2-128 deployed(v3) 100.00% 0.00% -100.00pp 通过
仲裁准确率 V2-128 deployed(v3) 81.12% 100.00% +18.88pp 通过
仲裁准确率 V2-128 deployed(v3) 81.12% 100.00% +18.88pp 通过
仲裁准确率 V2-128 deployed(v3) 81.12% 100.00% +18.88pp 通过
仲裁准确率 V2-128 deployed(v3) 81.12% 100.00% +18.88pp 通过
平均分数余量 V2-128 deployed(v3) -9.2989 40.1456 +49.4445 通过
平均分数余量 V2-128 deployed(v3) -9.2989 21.8925 +31.1914 通过
平均分数余量 V2-128 deployed(v3) -9.2989 18.4219 +27.7208 通过
平均分数余量 V2-128 deployed(v3) -9.2989 14.9592 +24.2581 通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 0.9698 0.9359 -0.0339 通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 0.9698 0.9776 +0.0078 通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 0.9698 1.2315 +0.2617 未通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 0.9698 1.2807 +0.3109 未通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 1031.1135 1068.4453 +37.3318 通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 1031.1135 1022.9307 -8.1828 通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 1031.1135 812.0051 -219.1084 未通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 1031.1135 780.7949 -250.3186 未通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 64355.2410 65927.7268 +1572.4858 通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 64355.2410 73335.6250 +8980.3841 通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 64355.2410 49115.9135 -15239.3274 未通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 64355.2410 62429.2793 -1925.9616 通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 185281.6861 262171.5170 +76889.8309 通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 185281.6861 273837.7940 +88556.1079 通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 185281.6861 141564.7326 -43716.9535 未通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 185281.6861 201612.9032 +16331.2171 通过
地址字节/记录 V2-128 deployed(v3) 512.0000 512.0000 +0.0000 通过
地址字节/记录 V2-128 deployed(v3) 512.0000 512.0000 +0.0000 通过
地址字节/记录 V2-128 deployed(v3) 512.0000 512.0000 +0.0000 通过
地址字节/记录 V2-128 deployed(v3) 512.0000 512.0000 +0.0000 通过

逐候选判定(全方位超越是对单个路由器说的)

路由器 通过 未通过 全方位超越 未通过的轴
REPLAY-128 v7 22 0 是 -
PROD-128 step20000 22 0 是 -
PROD-128 step10000 18 4 否 单查询延迟 ms (GPU), 路由 QPS (GPU, 单查询), 批量 QPS (GPU, batch=64), 批量 QPS (GPU, batch=256)
V2-128 v6 best 20 2 否 单查询延迟 ms (GPU), 路由 QPS (GPU, 单查询)

汇总 82 项通过 / 6 项未通过 / 0 项无数据