- 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
315 lines
6.5 KiB
JSON
315 lines
6.5 KiB
JSON
{
|
|
"package": "qwen3_5_4b_natural_memory_v2",
|
|
"facts": 20,
|
|
"phases": {
|
|
"A_distinct_facts": {
|
|
"records": 20,
|
|
"status_counts": {
|
|
"active": 20
|
|
},
|
|
"active": 20,
|
|
"superseded": 0,
|
|
"retracted": 0,
|
|
"attribute_status": {
|
|
"主管姓名": [
|
|
"active"
|
|
],
|
|
"入职年份": [
|
|
"active"
|
|
],
|
|
"出生城市": [
|
|
"active"
|
|
],
|
|
"办公城市": [
|
|
"active"
|
|
],
|
|
"午餐偏好": [
|
|
"active"
|
|
],
|
|
"咖啡口味": [
|
|
"active"
|
|
],
|
|
"团队名称": [
|
|
"active"
|
|
],
|
|
"工位楼层": [
|
|
"active"
|
|
],
|
|
"常住城市": [
|
|
"active"
|
|
],
|
|
"常用编辑器": [
|
|
"active"
|
|
],
|
|
"手机尾号": [
|
|
"active"
|
|
],
|
|
"档案标识": [
|
|
"active"
|
|
],
|
|
"紧急联系人姓氏": [
|
|
"active"
|
|
],
|
|
"起床时间": [
|
|
"active"
|
|
],
|
|
"运动习惯": [
|
|
"active"
|
|
],
|
|
"通勤方式": [
|
|
"active"
|
|
],
|
|
"邮箱域名": [
|
|
"active"
|
|
],
|
|
"阅读工具": [
|
|
"active"
|
|
],
|
|
"项目代号": [
|
|
"active"
|
|
],
|
|
"默认语言": [
|
|
"active"
|
|
]
|
|
}
|
|
},
|
|
"B_positive_control": {
|
|
"before": {
|
|
"records": 20,
|
|
"status_counts": {
|
|
"active": 20
|
|
},
|
|
"active": 20,
|
|
"superseded": 0,
|
|
"retracted": 0,
|
|
"attribute_status": {
|
|
"主管姓名": [
|
|
"active"
|
|
],
|
|
"入职年份": [
|
|
"active"
|
|
],
|
|
"出生城市": [
|
|
"active"
|
|
],
|
|
"办公城市": [
|
|
"active"
|
|
],
|
|
"午餐偏好": [
|
|
"active"
|
|
],
|
|
"咖啡口味": [
|
|
"active"
|
|
],
|
|
"团队名称": [
|
|
"active"
|
|
],
|
|
"工位楼层": [
|
|
"active"
|
|
],
|
|
"常住城市": [
|
|
"active"
|
|
],
|
|
"常用编辑器": [
|
|
"active"
|
|
],
|
|
"手机尾号": [
|
|
"active"
|
|
],
|
|
"档案标识": [
|
|
"active"
|
|
],
|
|
"紧急联系人姓氏": [
|
|
"active"
|
|
],
|
|
"起床时间": [
|
|
"active"
|
|
],
|
|
"运动习惯": [
|
|
"active"
|
|
],
|
|
"通勤方式": [
|
|
"active"
|
|
],
|
|
"邮箱域名": [
|
|
"active"
|
|
],
|
|
"阅读工具": [
|
|
"active"
|
|
],
|
|
"项目代号": [
|
|
"active"
|
|
],
|
|
"默认语言": [
|
|
"active"
|
|
]
|
|
}
|
|
},
|
|
"after": {
|
|
"records": 21,
|
|
"status_counts": {
|
|
"superseded": 1,
|
|
"active": 20
|
|
},
|
|
"active": 20,
|
|
"superseded": 1,
|
|
"retracted": 0,
|
|
"attribute_status": {
|
|
"主管姓名": [
|
|
"active"
|
|
],
|
|
"入职年份": [
|
|
"active"
|
|
],
|
|
"出生城市": [
|
|
"active"
|
|
],
|
|
"办公城市": [
|
|
"active"
|
|
],
|
|
"午餐偏好": [
|
|
"active"
|
|
],
|
|
"咖啡口味": [
|
|
"active"
|
|
],
|
|
"团队名称": [
|
|
"active"
|
|
],
|
|
"工位楼层": [
|
|
"active"
|
|
],
|
|
"常住城市": [
|
|
"active",
|
|
"superseded"
|
|
],
|
|
"常用编辑器": [
|
|
"active"
|
|
],
|
|
"手机尾号": [
|
|
"active"
|
|
],
|
|
"档案标识": [
|
|
"active"
|
|
],
|
|
"紧急联系人姓氏": [
|
|
"active"
|
|
],
|
|
"起床时间": [
|
|
"active"
|
|
],
|
|
"运动习惯": [
|
|
"active"
|
|
],
|
|
"通勤方式": [
|
|
"active"
|
|
],
|
|
"邮箱域名": [
|
|
"active"
|
|
],
|
|
"阅读工具": [
|
|
"active"
|
|
],
|
|
"项目代号": [
|
|
"active"
|
|
],
|
|
"默认语言": [
|
|
"active"
|
|
]
|
|
}
|
|
},
|
|
"常住城市_values": [
|
|
[
|
|
"VAL-0000000",
|
|
"superseded"
|
|
],
|
|
[
|
|
"VAL-9999999",
|
|
"active"
|
|
]
|
|
]
|
|
},
|
|
"C_retriever_absent": {
|
|
"records": 20,
|
|
"status_counts": {
|
|
"active": 20
|
|
},
|
|
"active": 20,
|
|
"superseded": 0,
|
|
"retracted": 0,
|
|
"attribute_status": {
|
|
"主管姓名": [
|
|
"active"
|
|
],
|
|
"入职年份": [
|
|
"active"
|
|
],
|
|
"出生城市": [
|
|
"active"
|
|
],
|
|
"办公城市": [
|
|
"active"
|
|
],
|
|
"午餐偏好": [
|
|
"active"
|
|
],
|
|
"咖啡口味": [
|
|
"active"
|
|
],
|
|
"团队名称": [
|
|
"active"
|
|
],
|
|
"工位楼层": [
|
|
"active"
|
|
],
|
|
"常住城市": [
|
|
"active"
|
|
],
|
|
"常用编辑器": [
|
|
"active"
|
|
],
|
|
"手机尾号": [
|
|
"active"
|
|
],
|
|
"档案标识": [
|
|
"active"
|
|
],
|
|
"紧急联系人姓氏": [
|
|
"active"
|
|
],
|
|
"起床时间": [
|
|
"active"
|
|
],
|
|
"运动习惯": [
|
|
"active"
|
|
],
|
|
"通勤方式": [
|
|
"active"
|
|
],
|
|
"邮箱域名": [
|
|
"active"
|
|
],
|
|
"阅读工具": [
|
|
"active"
|
|
],
|
|
"项目代号": [
|
|
"active"
|
|
],
|
|
"默认语言": [
|
|
"active"
|
|
]
|
|
}
|
|
}
|
|
},
|
|
"checks": {
|
|
"A_all_distinct_facts_active": true,
|
|
"A_no_retractions": true,
|
|
"B_new_value_is_active": true,
|
|
"B_old_version_superseded_not_destroyed": true,
|
|
"B_other_attributes_untouched": true,
|
|
"B_no_retractions": true,
|
|
"C_all_distinct_facts_active_without_retriever": true,
|
|
"C_no_retractions_without_retriever": true
|
|
},
|
|
"all_checks_passed": true
|
|
} |