- 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
265 lines
6.9 KiB
JSON
265 lines
6.9 KiB
JSON
{
|
|
"train": {
|
|
"episodes": 87155,
|
|
"families": {
|
|
"benchmark_qa": 128,
|
|
"native_memory": 512,
|
|
"memory_policy": 6515,
|
|
"mega_validation": 80000
|
|
},
|
|
"categories": {
|
|
"(none)": 7155,
|
|
"conflict_update": 8000,
|
|
"distractor_128": 8000,
|
|
"distractor_32": 8000,
|
|
"forget_correction": 8000,
|
|
"long_context": 8000,
|
|
"multi_hop": 8000,
|
|
"paraphrase": 8000,
|
|
"random_position": 8000,
|
|
"single_fact": 8000,
|
|
"unknown_abstention": 8000
|
|
},
|
|
"per_category": {
|
|
"(none)": {
|
|
"episodes": 7155,
|
|
"unknown": 319,
|
|
"candidates_32": 7155,
|
|
"positives_0": 319,
|
|
"hop_0": 319,
|
|
"candidate_slots": 228960,
|
|
"positive_total": 6836,
|
|
"answerable": 6836,
|
|
"positives_1": 6836,
|
|
"hop_1": 6836
|
|
},
|
|
"conflict_update": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
},
|
|
"distractor_128": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
},
|
|
"distractor_32": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
},
|
|
"forget_correction": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_2": 8000,
|
|
"hop_2": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 16000
|
|
},
|
|
"long_context": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
},
|
|
"multi_hop": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
},
|
|
"paraphrase": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
},
|
|
"random_position": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
},
|
|
"single_fact": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
},
|
|
"unknown_abstention": {
|
|
"episodes": 8000,
|
|
"answerable": 8000,
|
|
"candidates_32": 8000,
|
|
"positives_1": 8000,
|
|
"hop_1": 8000,
|
|
"candidate_slots": 256000,
|
|
"positive_total": 8000
|
|
}
|
|
},
|
|
"seconds": 10.4
|
|
},
|
|
"eval": {
|
|
"episodes": 21920,
|
|
"families": {
|
|
"benchmark_qa": 64,
|
|
"native_memory": 128,
|
|
"memory_policy": 1728,
|
|
"mega_validation": 20000
|
|
},
|
|
"categories": {
|
|
"(none)": 1920,
|
|
"conflict_update": 2000,
|
|
"distractor_128": 2000,
|
|
"distractor_32": 2000,
|
|
"forget_correction": 2000,
|
|
"long_context": 2000,
|
|
"multi_hop": 2000,
|
|
"paraphrase": 2000,
|
|
"random_position": 2000,
|
|
"single_fact": 2000,
|
|
"unknown_abstention": 2000
|
|
},
|
|
"per_category": {
|
|
"(none)": {
|
|
"episodes": 1920,
|
|
"unknown": 106,
|
|
"candidates_32": 1920,
|
|
"positives_0": 106,
|
|
"hop_0": 106,
|
|
"candidate_slots": 61440,
|
|
"positive_total": 1814,
|
|
"answerable": 1814,
|
|
"positives_1": 1814,
|
|
"hop_1": 1814
|
|
},
|
|
"conflict_update": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
},
|
|
"distractor_128": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
},
|
|
"distractor_32": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
},
|
|
"forget_correction": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_2": 2000,
|
|
"hop_2": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 4000
|
|
},
|
|
"long_context": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
},
|
|
"multi_hop": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
},
|
|
"paraphrase": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
},
|
|
"random_position": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
},
|
|
"single_fact": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
},
|
|
"unknown_abstention": {
|
|
"episodes": 2000,
|
|
"answerable": 2000,
|
|
"candidates_32": 2000,
|
|
"positives_1": 2000,
|
|
"hop_1": 2000,
|
|
"candidate_slots": 64000,
|
|
"positive_total": 2000
|
|
}
|
|
},
|
|
"seconds": 2.5
|
|
},
|
|
"leakage": {
|
|
"group_id_overlap": 0,
|
|
"eval_queries_seen_in_train": 80,
|
|
"eval_queries_total": 4296,
|
|
"eval_candidate_texts_seen_in_train": 11,
|
|
"eval_candidate_texts_total": 420819,
|
|
"query_overlap_rate": 0.0186219739292365,
|
|
"candidate_overlap_rate": 2.613950415736932e-05
|
|
}
|
|
} |