| metric | V2-512 best(step1000) | V2-512 final(step100k) | XL-512 best(now) | |---|---|---|---| | 参数 | 4,741,902 | 4,741,902 | 7,898,127 | | router_dim | 512 | 512 | 512 | | 每条记录地址字节 | 2048 | 2048 | 2048 | | **Top-1** | 0.6712 | 0.6063 | 0.6085 | | Recall@1 | 0.6712 | 0.6063 | 0.6085 | | Recall@3 | 0.8877 | 0.8501 | 0.9297 | | Recall@5 | 0.9493 | 0.9013 | 0.9733 | | MRR | 0.7908 | 0.7326 | 0.7598 | | nDCG@3 | 0.8000 | 0.7457 | 0.7928 | | 多跳证据全中(Top-3) | 0.8866 | 0.8490 | 0.9286 | | hop 正确率 | 0.9784 | 0.9969 | 0.9938 | | hop 欠预测率 | 0.0033 | 0.0033 | 0.0033 | | need F1 | 0.9892 | 0.9989 | 0.9962 | | need 召回 | 0.9978 | 0.9978 | 0.9978 | | **未知拒答率**(specificity) | 0.6604 | 1.0000 | 0.9057 | | **已知问题被误拒率** | 0.0022 | 0.0022 | 0.0022 | | 未知问题误读率 | 0.3396 | 0.0000 | 0.0943 | | 仲裁准确率 | 0.9794 | 0.9979 | 0.9928 | | 已知/未知 need 概率均值 | 0.985/0.254 | 0.998/0.015 | 0.996/0.135 | | 平均分数余量 | 0.5880 | 3.8837 | 6.7097 | | 单查询延迟 ms(GPU) | - | - | - | | 路由 QPS(GPU) | - | - | - |