wpy/natural-memory-nm21
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

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

Browse Source 
- 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
Download Patch File Download Diff File Expand all files Collapse all files
nm2_1_combo2_critical_e2e.json
+228
View File
@@ -0,0 +1,228 @@
{
"NM2.1 合体": {
"summary": {
"router": "NM2.1 合体",
"cases": 16,
"accuracy_pct": 75.0,
"wrong_attribute_pct": 12.5,
"read_pct": 93.75,
"cases_with_rule_hits": 6,
"stop_reasons": {
"evidence_found": 15,
"router_abstained": 1
},
"mean_records_selected": 6.0625,
"seconds": 85.6
},
"rows": [
{
"attribute": "常住城市",
"wanted": "CITY-A1B2C3",
"reply": "您的常住城市是 CITY-A1B2C3。",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 3.2870910186502984,
"records_selected": 8
},
{
"attribute": "出生城市",
"wanted": "CITY-D4E5F6",
"reply": "根据现有记忆，您的老家信息未知。已知事实包括：我的工位在 3 号楼；我的常住城市是 CITY-A1B2C3；我的出生城市是 CITY-D4E5F6；我的办公城市",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 2.2259970108983125,
"records_selected": 8
},
{
"attribute": "办公城市",
"wanted": "CITY-G7H8J9",
"reply": "您的办公城市是 CITY-G7H8J9。",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 20,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 8.520044145734458,
"records_selected": 8
},
{
"attribute": "档案代号",
"wanted": "CODE-K1L2M3",
"reply": "系统分配给您的字符是：user；紧急联系人姓氏；KIN-X1Y2Z3；团队名称；TEAM-D1E2F3；主管姓名；BOSS-X4Y5Z6；邮箱域名；",
"correct": false,
"wrong_attribute_code": true,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 3.698177008675168,
"records_selected": 8
},
{
"attribute": "常用编辑器",
"wanted": "EDIT-N4P5Q6",
"reply": "您写代码时使用的软件是 EDIT-N4P5Q6。\n常用编辑器：EDIT-N4P5Q6",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 3,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 8.515892826960929,
"records_selected": 3
},
{
"attribute": "默认语言",
"wanted": "LANG-R7S8T9",
"reply": "我的默认语言是 LANG-R7S8T9。user；默认语言；LANG-R7S8T9",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 2.400822959183151,
"records_selected": 8
},
{
"attribute": "通勤方式",
"wanted": "COMMU-U1V2W3",
"reply": "您的通勤方式是 COMMU-U1V2W3。",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 3.291031232642376,
"records_selected": 8
},
{
"attribute": "主管姓名",
"wanted": "BOSS-X4Y5Z6",
"reply": "这句话听起来带着几分疲惫、迷茫，或者是对现状的无力感。能感受到你此刻可能正经历一些难以独自应对的困境。\n\n“谁管我？”这个问题背后，或许藏着几个不同的需求，我想为你拆解一下",
"correct": false,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "router_abstained",
"need_memory": false,
"top_score": 0.0,
"records_selected": 0
},
{
"attribute": "工位楼层",
"wanted": "FLOOR-A7B8C9",
"reply": "您的工位楼层是 FLOOR-A7B8C9。",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 4.041911891364496,
"records_selected": 8
},
{
"attribute": "团队名称",
"wanted": "TEAM-D1E2F3",
"reply": "您属于 TEAM-D1E2F3 小组。",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 3.7999397155004493,
"records_selected": 8
},
{
"attribute": "邮箱域名",
"wanted": "MAIL-G4H5J6",
"reply": "关于别人发信需要写的后缀，当前记忆中没有相关信息，无法回答。已知信息仅包含：user；紧急联系人姓氏；KIN-X1Y2Z3。",
"correct": false,
"wrong_attribute_code": true,
"rule_address_hit": false,
"rule_lexical_hits": 1,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 8.513982709418043,
"records_selected": 1
},
{
"attribute": "手机尾号",
"wanted": "PHONE-K7L8M9",
"reply": "您的电话号码最后几位是 PHONE-K7L8M9。user；手机尾号；PHONE-K7L8M9",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 3.600087038612998,
"records_selected": 8
},
{
"attribute": "项目代号",
"wanted": "PROJ-N1P2Q3",
"reply": "关于您正在做的工程名称，当前长期记忆中没有相关记录，无法回答。",
"correct": false,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 2,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 8.514652462471835,
"records_selected": 2
},
{
"attribute": "入职年份",
"wanted": "YEAR-R4S5T6",
"reply": "您的入职年份是 YEAR-R4S5T6。",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 0,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 3.705354102880722,
"records_selected": 8
},
{
"attribute": "办公楼层",
"wanted": "DESK-U7V8W9",
"reply": "您的办公楼层是 DESK-U7V8W9。",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 3,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 8.524232190583632,
"records_selected": 3
},
{
"attribute": "紧急联系人姓氏",
"wanted": "KIN-X1Y2Z3",
"reply": "您的紧急联系人姓氏是 KIN-X1Y2Z3。",
"correct": true,
"wrong_attribute_code": false,
"rule_address_hit": false,
"rule_lexical_hits": 20,
"stop_reason": "evidence_found",
"need_memory": true,
"top_score": 8.522833668863877,
"records_selected": 8
}
]
}
}