Files
natural-memory-nm21/data/realistic_eval.manifest.json
WpyQwq 643e22ecb9 Natural Memory NM2.1: 记忆路由器分叉、数据集缺陷修复与全轴评测证据
- 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
2026-09-19 11:11:31 +08:00

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)"
}