Files
natural-memory-nm21/router_verdict_v6_128.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

12 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% 94.62% +53.49pp 通过
Top-1 正确率 V2-128 deployed(v3) 41.12% 94.37% +53.24pp 通过
Top-1 正确率 V2-128 deployed(v3) 41.12% 95.66% +54.54pp 通过
Top-1 正确率 V2-128 deployed(v3) 41.12% 95.19% +54.06pp 通过
Top-1 正确率 V2-128 deployed(v3) 41.12% 95.05% +53.92pp 通过
Top-1 正确率 V2-128 deployed(v3) 41.12% 95.78% +54.66pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 89.19% +52.16pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 88.82% +51.79pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 90.28% +53.25pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 89.64% +52.61pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 89.71% +52.68pp 通过
Recall@1 V2-128 deployed(v3) 37.03% 90.24% +53.21pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 97.81% +51.07pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 96.40% +49.67pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 98.75% +52.02pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 97.45% +50.72pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 98.84% +52.11pp 通过
Recall@3 V2-128 deployed(v3) 46.73% 97.59% +50.86pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 98.40% +49.49pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 97.40% +48.48pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 99.23% +50.31pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 98.20% +49.29pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 99.35% +50.44pp 通过
Recall@5 V2-128 deployed(v3) 48.91% 98.34% +49.43pp 通过
MRR V2-128 deployed(v3) 47.69% 96.49% +48.80pp 通过
MRR V2-128 deployed(v3) 47.69% 95.97% +48.28pp 通过
MRR V2-128 deployed(v3) 47.69% 97.37% +49.68pp 通过
MRR V2-128 deployed(v3) 47.69% 96.70% +49.01pp 通过
MRR V2-128 deployed(v3) 47.69% 97.12% +49.42pp 通过
MRR V2-128 deployed(v3) 47.69% 97.24% +49.55pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 96.49% +52.27pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 95.31% +51.09pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 97.51% +53.29pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 96.43% +52.21pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 97.24% +53.02pp 通过
nDCG@3 V2-128 deployed(v3) 44.22% 96.64% +52.42pp 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(Top-3) V2-128 deployed(v3) 43.43% 95.95% +52.53pp 通过
多跳证据全中(Top-3) V2-128 deployed(v3) 43.43% 98.70% +55.27pp 通过
多跳证据全中(Top-3) V2-128 deployed(v3) 43.43% 97.29% +53.86pp 通过
多跳证据全中(Top-3) V2-128 deployed(v3) 43.43% 98.64% +55.21pp 通过
多跳证据全中(Top-3) V2-128 deployed(v3) 43.43% 96.86% +53.44pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 99.00% +61.85pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 92.10% +54.95pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 99.10% +61.95pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 97.05% +59.90pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 96.35% +59.20pp 通过
多跳证据全中(仅多正例) V2-128 deployed(v3) 37.15% 87.05% +49.90pp 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% 99.95% +26.72pp 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
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% 99.95% +10.37pp 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% 99.98% +10.40pp 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
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% 99.59% +99.59pp 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% 99.81% +99.81pp 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
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% 0.41% -99.59pp 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% 0.19% -99.81pp 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% 99.92% +18.80pp 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% 99.96% +18.84pp 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 14.9592 +24.2581 通过
平均分数余量 V2-128 deployed(v3) -9.2989 51.3153 +60.6143 通过
平均分数余量 V2-128 deployed(v3) -9.2989 16.7980 +26.0969 通过
平均分数余量 V2-128 deployed(v3) -9.2989 200.3346 +209.6335 通过
平均分数余量 V2-128 deployed(v3) -9.2989 12.3693 +21.6682 通过
平均分数余量 V2-128 deployed(v3) -9.2989 45.5641 +54.8630 通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 1.1558 1.1679 +0.0121 通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 1.1558 1.1653 +0.0095 通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 1.1558 1.5783 +0.4225 未通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 1.1558 1.5959 +0.4401 未通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 1.1558 1.1623 +0.0064 通过
单查询延迟 ms (GPU) V2-128 deployed(v3) 1.1558 1.5855 +0.4297 未通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 865.1642 856.2377 -8.9265 通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 865.1642 858.1113 -7.0529 通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 865.1642 633.5730 -231.5911 未通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 865.1642 626.6057 -238.5585 未通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 865.1642 860.4001 -4.7641 通过
路由 QPS (GPU, 单查询) V2-128 deployed(v3) 865.1642 630.7159 -234.4483 未通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 54038.5362 54149.1810 +110.6448 通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 54038.5362 53471.4680 -567.0683 通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 54038.5362 38630.5471 -15407.9891 未通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 54038.5362 36843.4384 -17195.0978 未通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 54038.5362 53757.1186 -281.4176 通过
批量 QPS (GPU, batch=64) V2-128 deployed(v3) 54038.5362 39909.7043 -14128.8319 未通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 199554.1212 196829.2045 -2724.9167 通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 199554.1212 198025.9290 -1528.1922 通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 199554.1212 141500.5693 -58053.5519 未通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 199554.1212 134976.9590 -64577.1622 未通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 199554.1212 154085.0598 -45469.0615 未通过
批量 QPS (GPU, batch=256) V2-128 deployed(v3) 199554.1212 94647.2541 -104906.8671 未通过
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
地址字节/记录 V2-128 deployed(v3) 512.0000 2048.0000 +1536.0000 未通过
地址字节/记录 V2-128 deployed(v3) 512.0000 2048.0000 +1536.0000 未通过

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
V2-128 v6 best 22 0 是 -
V2-128 v6 final 22 0 是 -
XL-128 v6 best 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)
XL-128 v6 final 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)
V2-512 v6 best 20 2 否 批量 QPS (GPU, batch=256), 地址字节/记录
XL-512 v6 best 17 5 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), 地址字节/记录

汇总 117 项通过 / 15 项未通过 / 0 项无数据