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:
@@ -0,0 +1,59 @@
|
||||
{
|
||||
"原版NM2": {
|
||||
"A_cases": 110,
|
||||
"A_overall": 89.0909090909091,
|
||||
"A_answerable": 100.0,
|
||||
"A_unknown_refusal": 60.0,
|
||||
"A_known_false_refusal": 0.0,
|
||||
"B_accuracy": 43.75,
|
||||
"B_wrong_attribute": 18.75,
|
||||
"B_read": 56.25,
|
||||
"B_records": 3.0625,
|
||||
"C_answerable": 65.0,
|
||||
"C_wrong_attribute": 35.0,
|
||||
"C_unknown_leak": 75.0,
|
||||
"C_active_min": 23,
|
||||
"C_active_max": 24,
|
||||
"D_recalled": true,
|
||||
"D_answer_correct": true,
|
||||
"D_cleanup": true
|
||||
},
|
||||
"NM2.1": {
|
||||
"A_cases": 110,
|
||||
"A_overall": 88.18181818181819,
|
||||
"A_answerable": 100.0,
|
||||
"A_unknown_refusal": 56.666666666666664,
|
||||
"A_known_false_refusal": 0.0,
|
||||
"B_accuracy": 68.75,
|
||||
"B_wrong_attribute": 18.75,
|
||||
"B_read": 93.75,
|
||||
"B_records": 6.0625,
|
||||
"C_answerable": 65.0,
|
||||
"C_wrong_attribute": 35.0,
|
||||
"C_unknown_leak": 75.0,
|
||||
"C_active_min": 23,
|
||||
"C_active_max": 24,
|
||||
"D_recalled": true,
|
||||
"D_answer_correct": true,
|
||||
"D_cleanup": true
|
||||
},
|
||||
"NM2.1最终": {
|
||||
"A_cases": 110,
|
||||
"A_overall": 88.18181818181819,
|
||||
"A_answerable": 100.0,
|
||||
"A_unknown_refusal": 56.666666666666664,
|
||||
"A_known_false_refusal": 0.0,
|
||||
"B_accuracy": 68.75,
|
||||
"B_wrong_attribute": 18.75,
|
||||
"B_read": 93.75,
|
||||
"B_records": 6.0625,
|
||||
"C_answerable": 70.0,
|
||||
"C_wrong_attribute": 27.5,
|
||||
"C_unknown_leak": 0.0,
|
||||
"C_active_min": 23,
|
||||
"C_active_max": 24,
|
||||
"D_recalled": true,
|
||||
"D_answer_correct": true,
|
||||
"D_cleanup": true
|
||||
}
|
||||
}
|
||||
Reference in New Issue
Block a user