WpyQwq
643e22ecb9
- 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
35 lines
907 B
JSON
35 lines
907 B
JSON
{
|
|
"held_out_attributes": [
|
|
"主管姓名",
|
|
"办公城市",
|
|
"咖啡口味",
|
|
"常住城市",
|
|
"手机尾号",
|
|
"档案标识",
|
|
"紧急联系人姓氏",
|
|
"课程名称",
|
|
"通勤方式",
|
|
"邮箱域名",
|
|
"项目代号",
|
|
"默认语言"
|
|
],
|
|
"trained_on_attributes": 12,
|
|
"eval_held_out_queries": 138,
|
|
"eval_kept_queries": 112,
|
|
"max_prob_held_out": {
|
|
"p10": 0.1377,
|
|
"p50": 0.1971,
|
|
"p90": 0.3075
|
|
},
|
|
"max_prob_kept": {
|
|
"p10": 0.3099,
|
|
"p50": 0.4342,
|
|
"p90": 0.5683
|
|
},
|
|
"openset_auc_held_out_vs_kept": 0.976,
|
|
"closed_set_false_claim_pct": 100.0,
|
|
"reject_threshold_from_kept_p05": 0.2875,
|
|
"held_out_rejected_pct_at_threshold": 80.43,
|
|
"kept_retained_pct_at_threshold": 94.64,
|
|
"note": "held-out attributes never appear as training labels, so this is the open-set case a real user bank faces"
|
|
} |