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
13 lines
471 B
JSON
13 lines
471 B
JSON
{
|
|
"attribute_names": 24,
|
|
"episodes": 300,
|
|
"answerable": 250,
|
|
"abstention": 50,
|
|
"mean_present_attributes": 23.83,
|
|
"top1_attribute_accuracy_answerable_pct": 49.2,
|
|
"chance_pct": 4.2,
|
|
"best_score_answerable_p50": 0.7134,
|
|
"best_score_abstention_p50": 0.689,
|
|
"best_score_auc_covered_vs_uncovered": 0.656,
|
|
"note": "training-free: cosine between the frozen query key and the frozen encoding of each attribute NAME actually present in the bank"
|
|
} |