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
95 lines
2.7 KiB
JSON
95 lines
2.7 KiB
JSON
{
|
|
"package": "qwen3_5_4b_natural_memory_v2_1",
|
|
"eval_file": "data/realistic_v2/eval.jsonl",
|
|
"episodes": 24,
|
|
"per_category": {
|
|
"alias_paraphrase": {
|
|
"episodes": 3,
|
|
"candidates_written": 15,
|
|
"active_after_write": 15,
|
|
"write_survival_pct": 100.0,
|
|
"candidates_parsed_as_attribute_pct": 33.33,
|
|
"answerable_episodes": 3,
|
|
"target_survived": 3,
|
|
"target_survival_pct": 100.0
|
|
},
|
|
"long_fact": {
|
|
"episodes": 3,
|
|
"candidates_written": 12,
|
|
"active_after_write": 11,
|
|
"write_survival_pct": 91.67,
|
|
"candidates_parsed_as_attribute_pct": 0.0,
|
|
"answerable_episodes": 3,
|
|
"target_survived": 3,
|
|
"target_survival_pct": 100.0
|
|
},
|
|
"multi_entity": {
|
|
"episodes": 3,
|
|
"candidates_written": 9,
|
|
"active_after_write": 9,
|
|
"write_survival_pct": 100.0,
|
|
"candidates_parsed_as_attribute_pct": 0.0,
|
|
"answerable_episodes": 3,
|
|
"target_survived": 3,
|
|
"target_survival_pct": 100.0
|
|
},
|
|
"multi_hop": {
|
|
"episodes": 3,
|
|
"candidates_written": 15,
|
|
"active_after_write": 15,
|
|
"write_survival_pct": 100.0,
|
|
"candidates_parsed_as_attribute_pct": 0.0,
|
|
"answerable_episodes": 3,
|
|
"target_survived": 3,
|
|
"target_survival_pct": 100.0
|
|
},
|
|
"near_miss": {
|
|
"episodes": 3,
|
|
"candidates_written": 12,
|
|
"active_after_write": 12,
|
|
"write_survival_pct": 100.0,
|
|
"candidates_parsed_as_attribute_pct": 75.0,
|
|
"answerable_episodes": 3,
|
|
"target_survived": 3,
|
|
"target_survival_pct": 100.0
|
|
},
|
|
"noise_context": {
|
|
"episodes": 3,
|
|
"candidates_written": 27,
|
|
"active_after_write": 27,
|
|
"write_survival_pct": 100.0,
|
|
"candidates_parsed_as_attribute_pct": 11.11,
|
|
"answerable_episodes": 3,
|
|
"target_survived": 3,
|
|
"target_survival_pct": 100.0
|
|
},
|
|
"unknown_attribute": {
|
|
"episodes": 3,
|
|
"candidates_written": 15,
|
|
"active_after_write": 15,
|
|
"write_survival_pct": 100.0,
|
|
"candidates_parsed_as_attribute_pct": 100.0,
|
|
"answerable_episodes": 0,
|
|
"target_survived": 0,
|
|
"target_survival_pct": 0.0
|
|
},
|
|
"update_conflict": {
|
|
"episodes": 3,
|
|
"candidates_written": 15,
|
|
"active_after_write": 6,
|
|
"write_survival_pct": 40.0,
|
|
"candidates_parsed_as_attribute_pct": 80.0,
|
|
"answerable_episodes": 3,
|
|
"target_survived": 3,
|
|
"target_survival_pct": 100.0
|
|
}
|
|
},
|
|
"target_missing_examples": [],
|
|
"overall": {
|
|
"candidates_written": 120,
|
|
"active_after_write": 110,
|
|
"write_survival_pct": 91.67,
|
|
"answerable_episodes": 21,
|
|
"target_survival_pct": 100.0
|
|
}
|
|
} |