- 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
76 lines
1.7 KiB
JSON
76 lines
1.7 KiB
JSON
{
|
|
"generator": "make_realistic_memory_data.py",
|
|
"split": "eval",
|
|
"episodes": 200,
|
|
"per_category": 25,
|
|
"seed": 20260914,
|
|
"sha256": "b5e4d771e1bbd4d30048b7c545c2d77b68728f0f34506c5964374e8170d3b1bd",
|
|
"categories": {
|
|
"multi_entity": 25,
|
|
"alias_paraphrase": 25,
|
|
"update_conflict": 25,
|
|
"multi_hop": 25,
|
|
"near_miss": 25,
|
|
"noise_context": 25,
|
|
"long_fact": 25,
|
|
"unknown_attribute": 25
|
|
},
|
|
"attribute_families": [
|
|
"主库地址",
|
|
"仓库库位",
|
|
"代码仓库",
|
|
"供应商",
|
|
"保险到期",
|
|
"值班电话",
|
|
"值班表",
|
|
"发布窗口",
|
|
"合同编号",
|
|
"告警阈值",
|
|
"周会时间",
|
|
"培训周期"
|
|
],
|
|
"disjointness": {
|
|
"train_families": [
|
|
"备份周期",
|
|
"审批人",
|
|
"客户等级",
|
|
"对接端口",
|
|
"对账日",
|
|
"常驻机房",
|
|
"应急联系人",
|
|
"开源协议",
|
|
"接口版本号",
|
|
"日志保留",
|
|
"样机编号",
|
|
"灰度比例",
|
|
"监控面板",
|
|
"结算币种",
|
|
"续约提醒",
|
|
"缓存容量",
|
|
"证书到期日",
|
|
"质检标准",
|
|
"运输方式",
|
|
"返修地址",
|
|
"镜像标签",
|
|
"限流上限",
|
|
"预算上限",
|
|
"风速上限"
|
|
],
|
|
"eval_families": [
|
|
"主库地址",
|
|
"仓库库位",
|
|
"代码仓库",
|
|
"供应商",
|
|
"保险到期",
|
|
"值班电话",
|
|
"值班表",
|
|
"发布窗口",
|
|
"合同编号",
|
|
"告警阈值",
|
|
"周会时间",
|
|
"培训周期"
|
|
],
|
|
"overlap": []
|
|
},
|
|
"note": "eval families never appear in training, so eval results measure generalisation to unseen attribute wording"
|
|
} |