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
This commit is contained in:
WpyQwq
2026-09-19 11:11:31 +08:00
commit 643e22ecb9
484 changed files with 306821 additions and 0 deletions
+55
View File
@@ -0,0 +1,55 @@
# 拒答策略轴 · 全门槛曲线验证
基线:`V2-128 deployed(v3)`。门槛取值:0.30, 0.40, 0.50, 0.60, 0.70, 0.80。
全部为百分比。`通过` 表示该候选在每个门槛上都不劣于基线。
| 轴 | 门槛 0.30 | 门槛 0.40 | 门槛 0.50 | 门槛 0.60 | 门槛 0.70 | 门槛 0.80 | 全门槛通过 |
|---|---|---|---|---|---|---|---|
| V2-128 v6 final · need F1 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-128 v6 final · need 召回 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-128 v6 final · need 精确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-128 v6 final · 未知拒答率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-128 v6 final · 已知问题被误拒率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| V2-128 v6 final · 未知问题被误读率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| V2-128 v6 final · 仲裁准确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| REPLAY-128 v7 final · need F1 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| REPLAY-128 v7 final · need 召回 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| REPLAY-128 v7 final · need 精确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| REPLAY-128 v7 final · 未知拒答率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| REPLAY-128 v7 final · 已知问题被误拒率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| REPLAY-128 v7 final · 未知问题被误读率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| REPLAY-128 v7 final · 仲裁准确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-512 v6 final · need F1 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-512 v6 final · need 召回 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-512 v6 final · need 精确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-512 v6 final · 未知拒答率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| V2-512 v6 final · 已知问题被误拒率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| V2-512 v6 final · 未知问题被误读率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| V2-512 v6 final · 仲裁准确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-512 v6 final · need F1 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-512 v6 final · need 召回 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-512 v6 final · need 精确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-512 v6 final · 未知拒答率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-512 v6 final · 已知问题被误拒率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| XL-512 v6 final · 未知问题被误读率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| XL-512 v6 final · 仲裁准确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-128 v6 final · need F1 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-128 v6 final · need 召回 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-128 v6 final · need 精确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-128 v6 final · 未知拒答率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
| XL-128 v6 final · 已知问题被误拒率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| XL-128 v6 final · 未知问题被误读率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 是 |
| XL-128 v6 final · 仲裁准确率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 是 |
## 基线在各门槛上的值(作对照)
| 轴 | 门槛 0.30 | 门槛 0.40 | 门槛 0.50 | 门槛 0.60 | 门槛 0.70 | 门槛 0.80 |
|---|---|---|---|---|---|---|
| V2-128 deployed(v3) · need F1 | 89.67% | 89.67% | 89.58% | 89.30% | 89.19% | 89.01% |
| V2-128 deployed(v3) · need 召回 | 100.00% | 100.00% | 99.82% | 99.26% | 99.05% | 98.68% |
| V2-128 deployed(v3) · need 精确率 | 81.27% | 81.27% | 81.24% | 81.15% | 81.12% | 81.07% |
| V2-128 deployed(v3) · 未知拒答率 | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% | 0.00% |
| V2-128 deployed(v3) · 已知问题被误拒率 | 0.00% | 0.00% | 0.18% | 0.74% | 0.95% | 1.32% |
| V2-128 deployed(v3) · 未知问题被误读率 | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% | 100.00% |
| V2-128 deployed(v3) · 仲裁准确率 | 81.27% | 81.27% | 81.12% | 80.67% | 80.50% | 80.20% |