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
22 lines
633 B
JSON
22 lines
633 B
JSON
{
|
|
"package": "qwen3_5_4b_natural_memory_v2",
|
|
"router": "checkpoints\\router_replay_v7_v2_128\\memory_router_v2.pt",
|
|
"cases": 16,
|
|
"note": "blend 0.00 is the historical behaviour (packaged retriever only); 1.00 is the trained router only",
|
|
"results": {
|
|
"0.00": {
|
|
"router": "blend=0.00",
|
|
"cases": 16,
|
|
"accuracy_pct": 68.75,
|
|
"wrong_attribute_pct": 18.75,
|
|
"read_pct": 93.75,
|
|
"cases_with_rule_hits": 6,
|
|
"stop_reasons": {
|
|
"evidence_found": 15,
|
|
"router_abstained": 1
|
|
},
|
|
"mean_records_selected": 6.0625,
|
|
"seconds": 81.1
|
|
}
|
|
}
|
|
} |