- 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
59 lines
1.5 KiB
JSON
59 lines
1.5 KiB
JSON
{
|
|
"generator": "make_realistic_memory_eval.py",
|
|
"purpose": "realistic-shaped memory evaluation corpus (multi-entity, aliases, updates, multi-hop, near-miss, noise, long facts, unknowns)",
|
|
"per_category": 25,
|
|
"seed": 20260913,
|
|
"episodes": 200,
|
|
"sha256": "333a3c02c6b5fede753bea15d597de655287677ad5b08230da36039867238035",
|
|
"categories": {
|
|
"multi_entity": {
|
|
"episodes": 25,
|
|
"answerable": 25,
|
|
"unknown": 0,
|
|
"max_candidates": 3
|
|
},
|
|
"alias_paraphrase": {
|
|
"episodes": 25,
|
|
"answerable": 25,
|
|
"unknown": 0,
|
|
"max_candidates": 5
|
|
},
|
|
"update_conflict": {
|
|
"episodes": 25,
|
|
"answerable": 25,
|
|
"unknown": 0,
|
|
"max_candidates": 5
|
|
},
|
|
"multi_hop": {
|
|
"episodes": 25,
|
|
"answerable": 25,
|
|
"unknown": 0,
|
|
"max_candidates": 5
|
|
},
|
|
"near_miss": {
|
|
"episodes": 25,
|
|
"answerable": 25,
|
|
"unknown": 0,
|
|
"max_candidates": 4
|
|
},
|
|
"noise_context": {
|
|
"episodes": 25,
|
|
"answerable": 25,
|
|
"unknown": 0,
|
|
"max_candidates": 9
|
|
},
|
|
"long_fact": {
|
|
"episodes": 25,
|
|
"answerable": 25,
|
|
"unknown": 0,
|
|
"max_candidates": 4
|
|
},
|
|
"unknown_attribute": {
|
|
"episodes": 25,
|
|
"answerable": 0,
|
|
"unknown": 25,
|
|
"max_candidates": 5
|
|
}
|
|
},
|
|
"note": "episodes use the schema consumed by eval_end_to_end_memory (query / candidates[].text / positive_indices / metadata.acceptable)"
|
|
} |