Files
natural-memory-nm21/router_scorecard_v6_128.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

6132 lines
195 KiB
JSON

{
"V2-128 deployed(v3)": {
"checkpoint": "H:\\Memory\\dynamic_memory_lab\\checkpoints\\natural_memory_v2_qwen_router_entities\\memory_router_v2.pt",
"router": {
"kind": "router_v2",
"step": null,
"arch_config": {
"arch": "router_v2",
"hidden_size": 2560,
"router_dim": 128,
"num_heads": 8,
"max_hops": 3
},
"parameters": 2037774,
"checkpoint_bytes": 8157561
},
"storage": {
"address_bytes_per_record": 512,
"address_mb_per_1m_records": 512,
"checkpoint_bytes": 8157561
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 1.155849999122438,
"single_query_latency_ms_p50": 1.155849999122438,
"single_query_latency_ms_p95": 1.1732999992091209,
"queries_per_second": 865.1641655571528,
"batched_qps_64": 54038.53621622001,
"batched_qps_256": 199554.12120495233,
"batched_latency_ms_64": 0.01850531250511267,
"samples": 5,
"source": "bench_router_latency.py (interleaved round-robin medians)"
},
"cpu": {
"single_query_latency_ms_mean": 0.38441666644454625,
"single_query_latency_ms_p50": 0.36130000080447644,
"single_query_latency_ms_p95": 0.5763999979535583,
"queries_per_second": 2601.3440292507566,
"samples": 30,
"batched_qps_64": 12592.97154659022,
"batched_latency_ms_64": 0.07940937500734435,
"batched_qps_256": 13508.49346529746,
"batched_latency_ms_256": 0.07402749999982916
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.4112495789828225,
"recall1": 0.37027057370607386,
"recall3": 0.4673009992141013,
"recall5": 0.4891377568204783,
"recall8": 0.5160267205568654,
"evidence_count_in_top3": 0.5420455821264174,
"hit3": 0.5003368137420007,
"mrr": 0.476908110965675,
"ndcg3": 0.44220238482445295,
"mean_first_positive_rank": 14.38671831144044,
"all_evidence_in_top3": 0.43426518468620184,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.3715,
"hop_accuracy": 0.7322992700729927,
"hop_accuracy_answerable": 0.9010890310991355,
"hop_under_prediction": 0.046817110138093636,
"hop_over_prediction": 0.05209385876277085,
"mean_score_margin": -9.29892828391031,
"median_score_margin": -9.269989609718323,
"negative_margin_rate": 0.5887504210171774,
"hop_confusion": {
"true_0_pred_0": 0,
"true_0_pred_1": 4042,
"true_0_pred_2": 64,
"true_1_pred_0": 834,
"true_1_pred_1": 14052,
"true_1_pred_2": 928,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9762653562039442,
"unknown_mean_need_prob": 0.998855464207621,
"thr0.30": {
"need_precision": 0.8126824817518248,
"need_recall": 1.0,
"need_f1": 0.8966628076710123,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.8126824817518248,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 17814,
"tn": 0,
"fp": 4106,
"fn": 0
}
},
"thr0.40": {
"need_precision": 0.8126824817518248,
"need_recall": 1.0,
"need_f1": 0.8966628076710123,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.8126824817518248,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 17814,
"tn": 0,
"fp": 4106,
"fn": 0
}
},
"thr0.50": {
"need_precision": 0.8124086257309941,
"need_recall": 0.998203660042663,
"need_f1": 0.895773512669387,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.8112226277372263,
"known_question_refusal_rate": 0.001796339957336926,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 17782,
"tn": 0,
"fp": 4106,
"fn": 32
}
},
"thr0.60": {
"need_precision": 0.8115476409032495,
"need_recall": 0.9925900976759852,
"need_f1": 0.892985202767537,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.8066605839416059,
"known_question_refusal_rate": 0.00740990232401482,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 17682,
"tn": 0,
"fp": 4106,
"fn": 132
}
},
"thr0.70": {
"need_precision": 0.8112270700197692,
"need_recall": 0.9905130796003143,
"need_f1": 0.8919499557689878,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.8049726277372263,
"known_question_refusal_rate": 0.00948692039968564,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 17645,
"tn": 0,
"fp": 4106,
"fn": 169
}
},
"thr0.80": {
"need_precision": 0.8106525247867189,
"need_recall": 0.9868081284383069,
"need_f1": 0.8900984835059115,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.8019616788321168,
"known_question_refusal_rate": 0.01319187156169305,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 17579,
"tn": 0,
"fp": 4106,
"fn": 235
}
}
},
"by_family": {
"benchmark_qa": {
"episodes": 64,
"answerable_episodes": 0,
"unknown_episodes": 64,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 0,
"true_0_pred_2": 64
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.9999341862276196,
"thr0.50": {
"need_precision": 0.0,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 0,
"tn": 0,
"fp": 64,
"fn": 0
}
}
},
"mega_validation": {
"episodes": 20000,
"answerable_episodes": 16000,
"unknown_episodes": 4000,
"top1": 0.442375,
"recall1": 0.39675,
"recall3": 0.49153125,
"recall5": 0.50303125,
"recall8": 0.51234375,
"evidence_count_in_top3": 0.57475,
"hit3": 0.5283125,
"mrr": 0.500402578720693,
"ndcg3": 0.4692792057097287,
"mean_first_positive_rank": 14.9525,
"all_evidence_in_top3": 0.45475,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.3715,
"hop_accuracy": 0.74955,
"hop_accuracy_answerable": 0.9369375,
"hop_under_prediction": 0.0160625,
"hop_over_prediction": 0.047,
"mean_score_margin": -8.091850390457548,
"median_score_margin": -6.662378132343292,
"negative_margin_rate": 0.557625,
"hop_confusion": {
"true_0_pred_0": 0,
"true_0_pred_1": 4000,
"true_0_pred_2": 0,
"true_1_pred_0": 257,
"true_1_pred_1": 12991,
"true_1_pred_2": 752,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9821553228273988,
"unknown_mean_need_prob": 0.9988265322595835,
"thr0.50": {
"need_precision": 0.8,
"need_recall": 1.0,
"need_f1": 0.888888888888889,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.8,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 16000,
"tn": 0,
"fp": 4000,
"fn": 0
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.14528412419449327,
"recall1": 0.14528412419449327,
"recall3": 0.2677211482132396,
"recall5": 0.38488576449912126,
"recall8": 0.5653192735793791,
"evidence_count_in_top3": 0.2677211482132396,
"hit3": 0.2677211482132396,
"mrr": 0.2801716446629072,
"ndcg3": 0.21516992510461316,
"mean_first_positive_rank": 9.195664909197422,
"all_evidence_in_top3": 0.2677211482132396,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.5740740740740741,
"hop_accuracy_answerable": 0.5811364967779731,
"hop_under_prediction": 0.3380199179847686,
"hop_over_prediction": 0.08084358523725835,
"mean_score_margin": -19.359866383266713,
"median_score_margin": -18.159542322158813,
"negative_margin_rate": 0.8547158758055068,
"hop_confusion": {
"true_0_pred_0": 0,
"true_0_pred_1": 21,
"true_0_pred_2": 0,
"true_1_pred_0": 577,
"true_1_pred_1": 992,
"true_1_pred_2": 138
},
"known_mean_need_prob": 0.9195722371566764,
"unknown_mean_need_prob": 0.9999671209426153,
"thr0.50": {
"need_precision": 0.9876179245283019,
"need_recall": 0.981253661394259,
"need_f1": 0.9844255069056715,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.9693287037037037,
"known_question_refusal_rate": 0.018746338605741066,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 1675,
"tn": 0,
"fp": 21,
"fn": 32
}
}
},
"native_memory": {
"episodes": 128,
"answerable_episodes": 107,
"unknown_episodes": 21,
"top1": 0.0,
"recall1": 0.0,
"recall3": 0.028037383177570093,
"recall5": 0.07476635514018691,
"recall8": 0.2803738317757009,
"evidence_count_in_top3": 0.028037383177570093,
"hit3": 0.028037383177570093,
"mrr": 0.1023068389893556,
"ndcg3": 0.01524233414552764,
"mean_first_positive_rank": 12.598130841121495,
"all_evidence_in_top3": 0.028037383177570093,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.5390625,
"hop_accuracy_answerable": 0.6448598130841121,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.35514018691588783,
"mean_score_margin": -29.291684916086286,
"median_score_margin": -29.07153034210205,
"negative_margin_rate": 1.0,
"hop_confusion": {
"true_0_pred_0": 0,
"true_0_pred_1": 21,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 69,
"true_1_pred_2": 38
},
"known_mean_need_prob": 0.9999633771236812,
"unknown_mean_need_prob": 0.9999671209426153,
"thr0.50": {
"need_precision": 0.8359375,
"need_recall": 1.0,
"need_f1": 0.9106382978723404,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.8359375,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 107,
"tn": 0,
"fp": 21,
"fn": 0
}
}
}
},
"by_category": {
"conflict_update": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.796,
"recall1": 0.796,
"recall3": 0.9125,
"recall5": 0.919,
"recall8": 0.927,
"evidence_count_in_top3": 0.9125,
"hit3": 0.9125,
"mrr": 0.8582544951434297,
"ndcg3": 0.8678012294946459,
"mean_first_positive_rank": 2.8275,
"all_evidence_in_top3": 0.9125,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.4955,
"hop_accuracy_answerable": 0.4955,
"hop_under_prediction": 0.1285,
"hop_over_prediction": 0.376,
"mean_score_margin": 21.036133953779935,
"median_score_margin": 21.400100469589233,
"negative_margin_rate": 0.204,
"hop_confusion": {
"true_1_pred_0": 257,
"true_1_pred_1": 991,
"true_1_pred_2": 752
},
"known_mean_need_prob": 0.9764632563889026,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_128": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.231,
"recall1": 0.231,
"recall3": 0.2535,
"recall5": 0.255,
"recall8": 0.257,
"evidence_count_in_top3": 0.2535,
"hit3": 0.2535,
"mrr": 0.26603145346142226,
"ndcg3": 0.24486859507142916,
"mean_first_positive_rank": 23.789,
"all_evidence_in_top3": 0.2535,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -27.718347060799598,
"median_score_margin": -38.59446668624878,
"negative_margin_rate": 0.769,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9723067943453789,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_32": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.2255,
"recall1": 0.2255,
"recall3": 0.2455,
"recall5": 0.2475,
"recall8": 0.2485,
"evidence_count_in_top3": 0.2455,
"hit3": 0.2455,
"mrr": 0.2596338049221357,
"ndcg3": 0.23792220044107196,
"mean_first_positive_rank": 24.0695,
"all_evidence_in_top3": 0.2455,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -28.903205164782705,
"median_score_margin": -39.01692509651184,
"negative_margin_rate": 0.7745,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9724746745526791,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"forget_correction": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 0,
"true_0_pred_1": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.99774328109622,
"thr0.50": {
"need_precision": 0.0,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 0,
"tn": 0,
"fp": 2000,
"fn": 0
}
}
},
"long_context": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.5255,
"recall1": 0.5255,
"recall3": 0.6245,
"recall5": 0.6265,
"recall8": 0.6295,
"evidence_count_in_top3": 0.6245,
"hit3": 0.6245,
"mrr": 0.5855216777342396,
"ndcg3": 0.585867169546431,
"mean_first_positive_rank": 12.2085,
"all_evidence_in_top3": 0.6245,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -1.0039909677058458,
"median_score_margin": 2.929476022720337,
"negative_margin_rate": 0.4745,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9875293919146061,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"multi_hop": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.73,
"recall1": 0.365,
"recall3": 0.66575,
"recall5": 0.72725,
"recall8": 0.77375,
"evidence_count_in_top3": 1.3315,
"hit3": 0.96,
"mrr": 0.84532448592014,
"ndcg3": 0.6658792959224611,
"mean_first_positive_rank": 1.59,
"all_evidence_in_top3": 0.3715,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.3715,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 2.786984693795443,
"median_score_margin": 5.27826452255249,
"negative_margin_rate": 0.27,
"hop_confusion": {
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9993210740983486,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.443,
"recall1": 0.443,
"recall3": 0.5525,
"recall5": 0.5615,
"recall8": 0.5665,
"evidence_count_in_top3": 0.5525,
"hit3": 0.5525,
"mrr": 0.5118924660255056,
"ndcg3": 0.5092063534375025,
"mean_first_positive_rank": 13.8255,
"all_evidence_in_top3": 0.5525,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -2.995655853047967,
"median_score_margin": -5.5251545906066895,
"negative_margin_rate": 0.557,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9880377069413662,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"random_position": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.249,
"recall1": 0.249,
"recall3": 0.287,
"recall5": 0.293,
"recall8": 0.3005,
"evidence_count_in_top3": 0.287,
"hit3": 0.287,
"mrr": 0.2919631936667751,
"ndcg3": 0.27173149797678653,
"mean_first_positive_rank": 22.067,
"all_evidence_in_top3": 0.287,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -19.72415616109967,
"median_score_margin": -21.928287506103516,
"negative_margin_rate": 0.751,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9656237106621265,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"single_fact": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.339,
"recall1": 0.339,
"recall3": 0.391,
"recall5": 0.3945,
"recall8": 0.396,
"evidence_count_in_top3": 0.391,
"hit3": 0.391,
"mrr": 0.3845990528918967,
"ndcg3": 0.3709573037875013,
"mean_first_positive_rank": 19.243,
"all_evidence_in_top3": 0.391,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -8.212566563799978,
"median_score_margin": -16.937832549214363,
"negative_margin_rate": 0.661,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9954859737157822,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"unknown": {
"episodes": 1920,
"answerable_episodes": 1814,
"unknown_episodes": 106,
"top1": 0.13671444321940462,
"recall1": 0.13671444321940462,
"recall3": 0.2535832414553473,
"recall5": 0.36659316427783906,
"recall8": 0.5485115766262404,
"evidence_count_in_top3": 0.2535832414553473,
"hit3": 0.2535832414553473,
"mrr": 0.26968017045834825,
"ndcg3": 0.2033770627933551,
"mean_first_positive_rank": 9.396361631753033,
"all_evidence_in_top3": 0.2535832414553473,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.5526041666666667,
"hop_accuracy_answerable": 0.5848952590959207,
"hop_under_prediction": 0.31808158765159866,
"hop_over_prediction": 0.0970231532524807,
"mean_score_margin": -19.945701324287494,
"median_score_margin": -18.862584590911865,
"negative_margin_rate": 0.8632855567805954,
"hop_confusion": {
"true_0_pred_0": 0,
"true_0_pred_1": 42,
"true_0_pred_2": 64,
"true_1_pred_0": 577,
"true_1_pred_1": 1061,
"true_1_pred_2": 176
},
"known_mean_need_prob": 0.9243141621712682,
"unknown_mean_need_prob": 0.9999472358316746,
"thr0.50": {
"need_precision": 0.9438559322033898,
"need_recall": 0.9823594266813671,
"need_f1": 0.9627228525121555,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.928125,
"known_question_refusal_rate": 0.017640573318632856,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 1782,
"tn": 0,
"fp": 106,
"fn": 32
}
}
},
"unknown_abstention": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 0,
"true_0_pred_1": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.9999097834229469,
"thr0.50": {
"need_precision": 0.0,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 0.0,
"abstention_accuracy": 0.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 1.0,
"counts": {
"tp": 0,
"tn": 0,
"fp": 2000,
"fn": 0
}
}
}
}
},
"V2-128 v6 best": {
"checkpoint": "checkpoints\\router_v6_v2_128\\router_best.pt",
"router": {
"kind": "router_v2",
"step": 6000,
"arch_config": {
"arch": "router_v2",
"hidden_size": 2560,
"router_dim": 128,
"num_heads": 8,
"max_hops": 3
},
"parameters": 2037774,
"checkpoint_bytes": 24472973
},
"storage": {
"address_bytes_per_record": 512,
"address_mb_per_1m_records": 512,
"checkpoint_bytes": 24472973
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 1.167899999927613,
"single_query_latency_ms_p50": 1.167899999927613,
"single_query_latency_ms_p95": 1.196150000396301,
"queries_per_second": 856.2376916362535,
"batched_qps_64": 54149.18098951274,
"batched_qps_256": 196829.20449441153,
"batched_latency_ms_64": 0.018467500001406734,
"samples": 5,
"source": "bench_router_latency.py (interleaved round-robin medians)"
},
"cpu": {
"single_query_latency_ms_mean": 0.33882666709056747,
"single_query_latency_ms_p50": 0.338400001055561,
"single_query_latency_ms_p95": 0.4341000021668151,
"queries_per_second": 2951.3615577746205,
"samples": 30,
"batched_qps_64": 12676.831899779869,
"batched_latency_ms_64": 0.07888406250913249,
"batched_qps_256": 13789.017263119036,
"batched_latency_ms_256": 0.07252148437544292
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.946165936903559,
"recall1": 0.8919108566296171,
"recall3": 0.9780509711462895,
"recall5": 0.984001347254968,
"recall8": 0.9933198607836533,
"evidence_count_in_top3": 1.0897608622431796,
"hit3": 0.9786123273829572,
"mrr": 0.9648813461092199,
"ndcg3": 0.9649494672503449,
"mean_first_positive_rank": 1.1781744695183562,
"all_evidence_in_top3": 0.9774896149096216,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.99,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 14.959203617273067,
"median_score_margin": 14.255205273628235,
"negative_margin_rate": 0.053834063096441,
"hop_confusion": {
"true_0_pred_0": 4106,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 15814,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999601252354449,
"unknown_mean_need_prob": 0.0012111918483046415,
"thr0.30": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.40": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.60": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.70": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.80": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
}
},
"by_family": {
"benchmark_qa": {
"episodes": 64,
"answerable_episodes": 0,
"unknown_episodes": 64,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 64
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.010367806648901023,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 64,
"fp": 0,
"fn": 0
}
}
},
"mega_validation": {
"episodes": 20000,
"answerable_episodes": 16000,
"unknown_episodes": 4000,
"top1": 0.976875,
"recall1": 0.91646875,
"recall3": 0.999375,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.12375,
"hit3": 1.0,
"mrr": 0.9883645833333333,
"ndcg3": 0.9901670270848972,
"mean_first_positive_rank": 1.0235625,
"all_evidence_in_top3": 0.99875,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.99,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 16.402275828796906,
"median_score_margin": 15.20901894569397,
"negative_margin_rate": 0.023125,
"hop_confusion": {
"true_0_pred_0": 4000,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 14000,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999701014049351,
"unknown_mean_need_prob": 6.905754663642937e-05,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 16000,
"tn": 4000,
"fp": 0,
"fn": 0
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.6854130052724078,
"recall1": 0.6854130052724078,
"recall3": 0.786174575278266,
"recall5": 0.836555360281195,
"recall8": 0.9302870533099004,
"evidence_count_in_top3": 0.786174575278266,
"hit3": 0.786174575278266,
"mrr": 0.7618258791916341,
"ndcg3": 0.7445377808477717,
"mean_first_positive_rank": 2.570591681312244,
"all_evidence_in_top3": 0.786174575278266,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 2.378015720335124,
"median_score_margin": 3.452073097229004,
"negative_margin_rate": 0.3145869947275923,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1707
},
"known_mean_need_prob": 0.9998671145696143,
"unknown_mean_need_prob": 0.0960329504062732,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 1707,
"tn": 21,
"fp": 0,
"fn": 0
}
}
},
"native_memory": {
"episodes": 128,
"answerable_episodes": 107,
"unknown_episodes": 21,
"top1": 0.514018691588785,
"recall1": 0.514018691588785,
"recall3": 0.8504672897196262,
"recall5": 0.9439252336448598,
"recall8": 1.0,
"evidence_count_in_top3": 0.8504672897196262,
"hit3": 0.8504672897196262,
"mrr": 0.6927681352914997,
"ndcg3": 0.7103867694592852,
"mean_first_positive_rank": 2.0841121495327104,
"all_evidence_in_top3": 0.8504672897196262,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -0.11619492766456069,
"median_score_margin": 0.07718276977539062,
"negative_margin_rate": 0.48598130841121495,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 107
},
"known_mean_need_prob": 0.9999521859338351,
"unknown_mean_need_prob": 0.0960329504062732,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 107,
"tn": 21,
"fp": 0,
"fn": 0
}
}
}
},
"by_category": {
"conflict_update": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9885,
"recall1": 0.9885,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99425,
"ndcg3": 0.9957556921660717,
"mean_first_positive_rank": 1.0115,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 18.511014288797973,
"median_score_margin": 17.0071280002594,
"negative_margin_rate": 0.0115,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999475074112415,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_128": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9755,
"recall1": 0.9755,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.98775,
"ndcg3": 0.9909577789625007,
"mean_first_positive_rank": 1.0245,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 17.432058019019664,
"median_score_margin": 17.119813442230225,
"negative_margin_rate": 0.0245,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.999997846186161,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_32": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9805,
"recall1": 0.9805,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99025,
"ndcg3": 0.9928031301946435,
"mean_first_positive_rank": 1.0195,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 18.200293550987727,
"median_score_margin": 18.367231130599976,
"negative_margin_rate": 0.0195,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999978821873665,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"forget_correction": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.00013719209501925888,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
},
"long_context": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.98,
"recall1": 0.98,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9899166666666667,
"ndcg3": 0.9925531301946434,
"mean_first_positive_rank": 1.0205,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 17.648428340390325,
"median_score_margin": 17.865666389465332,
"negative_margin_rate": 0.02,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999990122318267,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"multi_hop": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9665,
"recall1": 0.48325,
"recall3": 0.995,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.99,
"hit3": 1.0,
"mrr": 0.983,
"ndcg3": 0.9775122185202447,
"mean_first_positive_rank": 1.035,
"all_evidence_in_top3": 0.99,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.99,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 10.340590946737677,
"median_score_margin": 12.604959726333618,
"negative_margin_rate": 0.0335,
"hop_confusion": {
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999316669106484,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9855,
"recall1": 0.9855,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99275,
"ndcg3": 0.9946484814267861,
"mean_first_positive_rank": 1.0145,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 16.270763264942914,
"median_score_margin": 14.272640228271484,
"negative_margin_rate": 0.0145,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.999909707903862,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"random_position": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.978,
"recall1": 0.978,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.989,
"ndcg3": 0.9918804545785721,
"mean_first_positive_rank": 1.022,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 17.229778190023264,
"median_score_margin": 17.07593348622322,
"negative_margin_rate": 0.022,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999774325489997,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"single_fact": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9605,
"recall1": 0.9605,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.98,
"ndcg3": 0.9852253306357154,
"mean_first_positive_rank": 1.041,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 15.585280029475689,
"median_score_margin": 15.148456335067749,
"negative_margin_rate": 0.0395,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.999999755859375,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"unknown": {
"episodes": 1920,
"answerable_episodes": 1814,
"unknown_episodes": 106,
"top1": 0.675303197353914,
"recall1": 0.675303197353914,
"recall3": 0.7899669239250275,
"recall5": 0.8428886438809261,
"recall8": 0.934399117971334,
"evidence_count_in_top3": 0.7899669239250275,
"hit3": 0.7899669239250275,
"mrr": 0.7577524621038092,
"ndcg3": 0.74252336066113,
"mean_first_positive_rank": 2.541896361631753,
"all_evidence_in_top3": 0.7899669239250275,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 2.2308930415391117,
"median_score_margin": 3.026643395423889,
"negative_margin_rate": 0.324696802646086,
"hop_confusion": {
"true_0_pred_0": 106,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1814
},
"known_mean_need_prob": 0.9998721325607783,
"unknown_mean_need_prob": 0.04431059945842585,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 1814,
"tn": 106,
"fp": 0,
"fn": 0
}
}
},
"unknown_abstention": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 9.229982535998715e-07,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
}
}
},
"V2-128 v6 final": {
"checkpoint": "checkpoints\\router_v6_v2_128\\memory_router_v2.pt",
"router": {
"kind": "router_v2",
"step": null,
"arch_config": {
"arch": "router_v2",
"hidden_size": 2560,
"router_dim": 128,
"num_heads": 8,
"max_hops": 3
},
"parameters": 2037774,
"checkpoint_bytes": 8156921
},
"storage": {
"address_bytes_per_record": 512,
"address_mb_per_1m_records": 512,
"checkpoint_bytes": 8156921
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 1.165349998700549,
"single_query_latency_ms_p50": 1.165349998700549,
"single_query_latency_ms_p95": 1.1786999984906288,
"queries_per_second": 858.1112979920828,
"batched_qps_64": 53471.46796235134,
"batched_qps_256": 198025.9290094514,
"batched_latency_ms_64": 0.018701562498790736,
"samples": 5,
"source": "bench_router_latency.py (interleaved round-robin medians)"
},
"cpu": {
"single_query_latency_ms_mean": 0.3544133333586312,
"single_query_latency_ms_p50": 0.3417000007175375,
"single_query_latency_ms_p95": 0.4805999997188337,
"queries_per_second": 2821.564275032788,
"samples": 30,
"batched_qps_64": 13137.043999093688,
"batched_latency_ms_64": 0.07612062501038963,
"batched_qps_256": 13174.159298667208,
"batched_latency_ms_256": 0.07590617187247517
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.9436959694622207,
"recall1": 0.8882059054676097,
"recall3": 0.9639609296059279,
"recall5": 0.973981138430448,
"recall8": 0.9864151790726395,
"evidence_count_in_top3": 1.0717974626698104,
"hit3": 0.9683956438756035,
"mrr": 0.9597054333626427,
"ndcg3": 0.9530820526370456,
"mean_first_positive_rank": 1.2573818345121814,
"all_evidence_in_top3": 0.9595262153362524,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.921,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 51.315328869764656,
"median_score_margin": 56.234654903411865,
"negative_margin_rate": 0.05630403053777928,
"hop_confusion": {
"true_0_pred_0": 4106,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 15814,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999984274064,
"unknown_mean_need_prob": 4.836221294362452e-06,
"thr0.30": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.40": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.60": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.70": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.80": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
}
},
"by_family": {
"benchmark_qa": {
"episodes": 64,
"answerable_episodes": 0,
"unknown_episodes": 64,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 64
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 4.8406160974828055e-06,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 64,
"fp": 0,
"fn": 0
}
}
},
"mega_validation": {
"episodes": 20000,
"answerable_episodes": 16000,
"unknown_episodes": 4000,
"top1": 0.9838125,
"recall1": 0.92203125,
"recall3": 0.9950625,
"recall5": 0.99978125,
"recall8": 0.999875,
"evidence_count_in_top3": 1.115125,
"hit3": 1.0,
"mrr": 0.991875,
"ndcg3": 0.9876817889701721,
"mean_first_positive_rank": 1.016375,
"all_evidence_in_top3": 0.990125,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.921,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 57.110110210259094,
"median_score_margin": 59.90444374084473,
"negative_margin_rate": 0.0161875,
"hop_confusion": {
"true_0_pred_0": 4000,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 14000,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999999701977,
"unknown_mean_need_prob": 6.379481702447893e-11,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 16000,
"tn": 4000,
"fp": 0,
"fn": 0
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.5940246045694201,
"recall1": 0.5940246045694201,
"recall3": 0.6842413591095489,
"recall5": 0.7352079671939075,
"recall8": 0.8611599297012302,
"evidence_count_in_top3": 0.6842413591095489,
"hit3": 0.6842413591095489,
"mrr": 0.6760273585431921,
"ndcg3": 0.6466497456649108,
"mean_first_positive_rank": 3.4463971880492092,
"all_evidence_in_top3": 0.6842413591095489,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 0.27154229221702947,
"median_score_margin": 7.975786209106445,
"negative_margin_rate": 0.40597539543058,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1707
},
"known_mean_need_prob": 0.9999999838679872,
"unknown_mean_need_prob": 0.0004654159529796483,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 1707,
"tn": 21,
"fp": 0,
"fn": 0
}
}
},
"native_memory": {
"episodes": 128,
"answerable_episodes": 107,
"unknown_episodes": 21,
"top1": 0.5233644859813084,
"recall1": 0.5233644859813084,
"recall3": 0.7757009345794392,
"recall5": 0.9252336448598131,
"recall8": 0.9719626168224299,
"evidence_count_in_top3": 0.7757009345794392,
"hit3": 0.7757009345794392,
"mrr": 0.674896157840083,
"ndcg3": 0.6678873486315127,
"mean_first_positive_rank": 2.3738317757009346,
"all_evidence_in_top3": 0.7757009345794392,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -0.8786688875929217,
"median_score_margin": 0.4756507873535156,
"negative_margin_rate": 0.4766355140186916,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 107
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": 0.0004654159529796483,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 107,
"tn": 21,
"fp": 0,
"fn": 0
}
}
}
},
"by_category": {
"conflict_update": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.993,
"recall1": 0.993,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9965,
"ndcg3": 0.9974165082750002,
"mean_first_positive_rank": 1.007,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 54.4210530708991,
"median_score_margin": 56.92102670669556,
"negative_margin_rate": 0.007,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_128": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.975,
"recall1": 0.975,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9875,
"ndcg3": 0.9907732438392864,
"mean_first_positive_rank": 1.025,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 64.29146421498805,
"median_score_margin": 73.40720510482788,
"negative_margin_rate": 0.025,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_32": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9845,
"recall1": 0.9845,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99225,
"ndcg3": 0.9942794111803576,
"mean_first_positive_rank": 1.0155,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 65.52974343589041,
"median_score_margin": 73.97303533554077,
"negative_margin_rate": 0.0155,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"forget_correction": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 1.275862421441728e-10,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
},
"long_context": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9795,
"recall1": 0.9795,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9896666666666666,
"ndcg3": 0.9923685950714292,
"mean_first_positive_rank": 1.021,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 60.150181606799364,
"median_score_margin": 68.01980018615723,
"negative_margin_rate": 0.0205,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"multi_hop": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9885,
"recall1": 0.49425,
"recall3": 0.9605,
"recall5": 0.99825,
"recall8": 0.999,
"evidence_count_in_top3": 1.921,
"hit3": 1.0,
"mrr": 0.99425,
"ndcg3": 0.9452009954703022,
"mean_first_positive_rank": 1.0115,
"all_evidence_in_top3": 0.921,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.921,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 55.6439196386747,
"median_score_margin": 62.50904202461243,
"negative_margin_rate": 0.0115,
"hop_confusion": {
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999997615814,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.995,
"recall1": 0.995,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9975,
"ndcg3": 0.9981546487678573,
"mean_first_positive_rank": 1.005,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 48.13904330630228,
"median_score_margin": 49.25835180282593,
"negative_margin_rate": 0.005,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"random_position": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9805,
"recall1": 0.9805,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9901666666666666,
"ndcg3": 0.9927376653178577,
"mean_first_positive_rank": 1.02,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 50.85701001610607,
"median_score_margin": 54.9089093208313,
"negative_margin_rate": 0.0195,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"single_fact": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9745,
"recall1": 0.9745,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9871666666666666,
"ndcg3": 0.9905232438392865,
"mean_first_positive_rank": 1.026,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 57.84846639241278,
"median_score_margin": 63.093711376190186,
"negative_margin_rate": 0.0255,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"unknown": {
"episodes": 1920,
"answerable_episodes": 1814,
"unknown_episodes": 106,
"top1": 0.5898566703417861,
"recall1": 0.5898566703417861,
"recall3": 0.6896361631753032,
"recall5": 0.7464167585446527,
"recall8": 0.8676957001102535,
"evidence_count_in_top3": 0.6896361631753032,
"hit3": 0.6896361631753032,
"mrr": 0.6759606339151697,
"ndcg3": 0.6479024598421029,
"mean_first_positive_rank": 3.383131201764057,
"all_evidence_in_top3": 0.6896361631753032,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 0.20369631854576997,
"median_score_margin": 6.827788352966309,
"negative_margin_rate": 0.4101433296582139,
"hop_confusion": {
"true_0_pred_0": 106,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1814
},
"known_mean_need_prob": 0.9999999848195447,
"unknown_mean_need_prob": 0.00018733273071117103,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 1814,
"tn": 106,
"fp": 0,
"fn": 0
}
}
},
"unknown_abstention": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 3.3919047850488803e-15,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
}
}
},
"XL-128 v6 best": {
"checkpoint": "checkpoints\\router_v6_xl128\\router_best.pt",
"router": {
"kind": "router_xl",
"step": 1000,
"arch_config": {
"arch": "router_xl",
"arch_version": 1,
"hidden_size": 2560,
"router_dim": 128,
"num_heads": 8,
"max_hops": 3,
"encoder_layers": 2,
"encoder_hidden": 512,
"pair_blocks": 1,
"pair_hidden": 512,
"pair_expansion": 2,
"pair_dropout": 0.05,
"use_interaction": true,
"policy_layers": 2,
"policy_hidden": 512,
"policy_dropout": 0.0,
"learnable_cosine_scale": true
},
"parameters": 6714639,
"checkpoint_bytes": 81666921
},
"storage": {
"address_bytes_per_record": 512,
"address_mb_per_1m_records": 512,
"checkpoint_bytes": 81666921
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 1.578349998453632,
"single_query_latency_ms_p50": 1.578349998453632,
"single_query_latency_ms_p95": 1.6234499998972751,
"queries_per_second": 633.5730357523598,
"batched_qps_64": 38630.547104467114,
"batched_qps_256": 141500.569342867,
"batched_latency_ms_64": 0.02588625000043976,
"samples": 5,
"source": "bench_router_latency.py (interleaved round-robin medians)"
},
"cpu": {
"single_query_latency_ms_mean": 1.4055266667128308,
"single_query_latency_ms_p50": 1.3804000009258743,
"single_query_latency_ms_p95": 1.7585999994480517,
"queries_per_second": 711.4770738137224,
"samples": 30,
"batched_qps_64": 2431.7772491816463,
"batched_latency_ms_64": 0.4112218750037755,
"batched_qps_256": 2460.1366567422874,
"batched_latency_ms_256": 0.4064814843758313
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.9566071629055799,
"recall1": 0.9027730998091389,
"recall3": 0.9874817559223082,
"recall5": 0.9922532839339845,
"recall8": 0.9991018300213316,
"evidence_count_in_top3": 1.0992477826428653,
"hit3": 0.9879869765353093,
"mrr": 0.9737093504776227,
"ndcg3": 0.975110976145454,
"mean_first_positive_rank": 1.098574155158864,
"all_evidence_in_top3": 0.9869765353093073,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.991,
"hop_accuracy": 0.9995437956204379,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 16.798009918724365,
"median_score_margin": 17.34114497900009,
"negative_margin_rate": 0.04339283709442012,
"hop_confusion": {
"true_0_pred_0": 4096,
"true_0_pred_1": 10,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 15814,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9900794763436311,
"unknown_mean_need_prob": 0.036242366845311694,
"thr0.30": {
"need_precision": 0.9964201812283253,
"need_recall": 1.0,
"need_f1": 0.9982068810938025,
"specificity_unknown_refusal": 0.9844130540672187,
"abstention_accuracy": 0.997080291970803,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.015586945932781296,
"counts": {
"tp": 17814,
"tn": 4042,
"fp": 64,
"fn": 0
}
},
"thr0.40": {
"need_precision": 0.998150949739452,
"need_recall": 1.0,
"need_f1": 0.9990746193320434,
"specificity_unknown_refusal": 0.9919629810034096,
"abstention_accuracy": 0.9984945255474452,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.008037018996590355,
"counts": {
"tp": 17814,
"tn": 4073,
"fp": 33,
"fn": 0
}
},
"thr0.50": {
"need_precision": 0.9990466042285907,
"need_recall": 1.0,
"need_f1": 0.9995230747650443,
"specificity_unknown_refusal": 0.995859717486605,
"abstention_accuracy": 0.9992244525547446,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.004140282513395032,
"counts": {
"tp": 17814,
"tn": 4089,
"fp": 17,
"fn": 0
}
},
"thr0.60": {
"need_precision": 0.999326825984517,
"need_recall": 1.0,
"need_f1": 0.9996632996632996,
"specificity_unknown_refusal": 0.9970774476376035,
"abstention_accuracy": 0.9994525547445255,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0029225523623964927,
"counts": {
"tp": 17814,
"tn": 4094,
"fp": 12,
"fn": 0
}
},
"thr0.70": {
"need_precision": 0.9996632996632997,
"need_recall": 1.0,
"need_f1": 0.9998316214850985,
"specificity_unknown_refusal": 0.9985387238188017,
"abstention_accuracy": 0.9997262773722628,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0014612761811982464,
"counts": {
"tp": 17814,
"tn": 4100,
"fp": 6,
"fn": 0
}
},
"thr0.80": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
}
},
"by_family": {
"benchmark_qa": {
"episodes": 64,
"answerable_episodes": 0,
"unknown_episodes": 64,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 64
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.3642658165190369,
"thr0.50": {
"need_precision": 0.0,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 0.953125,
"abstention_accuracy": 0.953125,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.046875,
"counts": {
"tp": 0,
"tn": 61,
"fp": 3,
"fn": 0
}
}
},
"mega_validation": {
"episodes": 20000,
"answerable_episodes": 16000,
"unknown_episodes": 4000,
"top1": 0.9765625,
"recall1": 0.916625,
"recall3": 0.9994375,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.123875,
"hit3": 1.0,
"mrr": 0.9882291666666666,
"ndcg3": 0.9904076605853642,
"mean_first_positive_rank": 1.02375,
"all_evidence_in_top3": 0.998875,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.991,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 17.91327105785167,
"median_score_margin": 18.24349021911621,
"negative_margin_rate": 0.0234375,
"hop_confusion": {
"true_0_pred_0": 4000,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 14000,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9914327386058867,
"unknown_mean_need_prob": 0.027471872764872386,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 16000,
"tn": 4000,
"fp": 0,
"fn": 0
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.8060925600468658,
"recall1": 0.8060925600468658,
"recall3": 0.8840070298769771,
"recall5": 0.9197422378441711,
"recall8": 0.9906268306971294,
"evidence_count_in_top3": 0.8840070298769771,
"hit3": 0.8840070298769771,
"mrr": 0.8601396423013294,
"ndcg3": 0.8521063487572197,
"mean_first_positive_rank": 1.7311072056239016,
"all_evidence_in_top3": 0.8840070298769771,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.9971064814814815,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 7.396975287374308,
"median_score_margin": 7.794462203979492,
"negative_margin_rate": 0.19390743995313414,
"hop_confusion": {
"true_0_pred_0": 16,
"true_0_pred_1": 5,
"true_1_pred_0": 0,
"true_1_pred_1": 1707
},
"known_mean_need_prob": 0.98037854359034,
"unknown_mean_need_prob": 0.37168226071766447,
"thr0.50": {
"need_precision": 0.9959159859976663,
"need_recall": 1.0,
"need_f1": 0.997953814674072,
"specificity_unknown_refusal": 0.6666666666666666,
"abstention_accuracy": 0.9959490740740741,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.3333333333333333,
"counts": {
"tp": 1707,
"tn": 14,
"fp": 7,
"fn": 0
}
}
},
"native_memory": {
"episodes": 128,
"answerable_episodes": 107,
"unknown_episodes": 21,
"top1": 0.37383177570093457,
"recall1": 0.37383177570093457,
"recall3": 0.8504672897196262,
"recall5": 0.9906542056074766,
"recall8": 1.0,
"evidence_count_in_top3": 0.8504672897196262,
"hit3": 0.8504672897196262,
"mrr": 0.6143302180685358,
"ndcg3": 0.6500824519693008,
"mean_first_positive_rank": 2.196261682242991,
"all_evidence_in_top3": 0.8504672897196262,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.9609375,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 0.007242532534019969,
"median_score_margin": -0.2684917449951172,
"negative_margin_rate": 0.6261682242990654,
"hop_confusion": {
"true_0_pred_0": 16,
"true_0_pred_1": 5,
"true_1_pred_0": 0,
"true_1_pred_1": 107
},
"known_mean_need_prob": 0.9424841119864277,
"unknown_mean_need_prob": 0.37168226071766447,
"thr0.50": {
"need_precision": 0.9385964912280702,
"need_recall": 1.0,
"need_f1": 0.9683257918552036,
"specificity_unknown_refusal": 0.6666666666666666,
"abstention_accuracy": 0.9453125,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.3333333333333333,
"counts": {
"tp": 107,
"tn": 14,
"fp": 7,
"fn": 0
}
}
}
},
"by_category": {
"conflict_update": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9785,
"recall1": 0.9785,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.98925,
"ndcg3": 0.9920649897017864,
"mean_first_positive_rank": 1.0215,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 19.512430111046882,
"median_score_margin": 19.13399839401245,
"negative_margin_rate": 0.0215,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9832325124442577,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_128": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.978,
"recall1": 0.978,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.989,
"ndcg3": 0.9918804545785721,
"mean_first_positive_rank": 1.022,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 18.416588767559734,
"median_score_margin": 18.676148295402527,
"negative_margin_rate": 0.022,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9980810531675816,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_32": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.985,
"recall1": 0.985,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9924166666666666,
"ndcg3": 0.9943984814267861,
"mean_first_positive_rank": 1.0155,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 18.82138747824775,
"median_score_margin": 18.756187587976456,
"negative_margin_rate": 0.015,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9980952816009522,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"forget_correction": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.05116949350852519,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
},
"long_context": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9765,
"recall1": 0.9765,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.98825,
"ndcg3": 0.9913268492089293,
"mean_first_positive_rank": 1.0235,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 15.71517678581737,
"median_score_margin": 16.085471630096436,
"negative_margin_rate": 0.0235,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9980619028508663,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"multi_hop": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.959,
"recall1": 0.4795,
"recall3": 0.9955,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.991,
"hit3": 1.0,
"mrr": 0.9793333333333334,
"ndcg3": 0.9775264704150523,
"mean_first_positive_rank": 1.042,
"all_evidence_in_top3": 0.991,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.991,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 17.043865082176403,
"median_score_margin": 20.163950562477112,
"negative_margin_rate": 0.041,
"hop_confusion": {
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9835184065401554,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.987,
"recall1": 0.987,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9935,
"ndcg3": 0.995202086796429,
"mean_first_positive_rank": 1.013,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 18.87393392176647,
"median_score_margin": 18.269309908151627,
"negative_margin_rate": 0.013,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9775506886839866,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"random_position": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.977,
"recall1": 0.977,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9885,
"ndcg3": 0.9915113843321435,
"mean_first_positive_rank": 1.023,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 17.55705898122629,
"median_score_margin": 17.530375365167856,
"negative_margin_rate": 0.023,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9958070812821388,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"single_fact": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9715,
"recall1": 0.9715,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9855833333333334,
"ndcg3": 0.9893505682232151,
"mean_first_positive_rank": 1.0295,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 17.365727334972473,
"median_score_margin": 16.746063470840454,
"negative_margin_rate": 0.0285,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9971149822771549,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"unknown": {
"episodes": 1920,
"answerable_episodes": 1814,
"unknown_episodes": 106,
"top1": 0.7805953693495039,
"recall1": 0.7805953693495039,
"recall3": 0.8820286659316428,
"recall5": 0.9239250275633958,
"recall8": 0.9911797133406836,
"evidence_count_in_top3": 0.8820286659316428,
"hit3": 0.8820286659316428,
"mrr": 0.8456404094496708,
"ndcg3": 0.8401898344483404,
"mean_first_positive_rank": 1.7585446527012127,
"all_evidence_in_top3": 0.8820286659316428,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.9947916666666666,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 6.961086971625735,
"median_score_margin": 7.221376895904541,
"negative_margin_rate": 0.21940463065049615,
"hop_confusion": {
"true_0_pred_0": 96,
"true_0_pred_1": 10,
"true_1_pred_0": 0,
"true_1_pred_1": 1814
},
"known_mean_need_prob": 0.9781433152653022,
"unknown_mean_need_prob": 0.3672044076166063,
"thr0.50": {
"need_precision": 0.9907154560349536,
"need_recall": 1.0,
"need_f1": 0.9953360768175583,
"specificity_unknown_refusal": 0.839622641509434,
"abstention_accuracy": 0.9911458333333333,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.16037735849056603,
"counts": {
"tp": 1814,
"tn": 89,
"fp": 17,
"fn": 0
}
}
},
"unknown_abstention": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.0037742520212195813,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
}
}
},
"XL-128 v6 final": {
"checkpoint": "checkpoints\\router_v6_xl128\\memory_router_xl.pt",
"router": {
"kind": "router_xl",
"step": null,
"arch_config": {
"arch": "router_xl",
"arch_version": 1,
"hidden_size": 2560,
"router_dim": 128,
"num_heads": 8,
"max_hops": 3,
"encoder_layers": 2,
"encoder_hidden": 512,
"pair_blocks": 1,
"pair_hidden": 512,
"pair_expansion": 2,
"pair_dropout": 0.05,
"use_interaction": true,
"policy_layers": 2,
"policy_hidden": 512,
"policy_dropout": 0.0,
"learnable_cosine_scale": true
},
"parameters": 6714639,
"checkpoint_bytes": 27922475
},
"storage": {
"address_bytes_per_record": 512,
"address_mb_per_1m_records": 512,
"checkpoint_bytes": 27922475
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 1.5958999993017642,
"single_query_latency_ms_p50": 1.5958999993017642,
"single_query_latency_ms_p95": 1.6187000001082197,
"queries_per_second": 626.6056773215856,
"batched_qps_64": 36843.43841344889,
"batched_qps_256": 134976.95897715187,
"batched_latency_ms_64": 0.027141875000324944,
"samples": 5,
"source": "bench_router_latency.py (interleaved round-robin medians)"
},
"cpu": {
"single_query_latency_ms_mean": 1.4327199997448286,
"single_query_latency_ms_p50": 1.4391999975487124,
"single_query_latency_ms_p95": 1.6943999980867375,
"queries_per_second": 697.9730862821087,
"samples": 30,
"batched_qps_64": 2507.9254362757065,
"batched_latency_ms_64": 0.3987359374946209,
"batched_qps_256": 2349.3282205997198,
"batched_latency_ms_256": 0.4256535937514627
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.9518917705175705,
"recall1": 0.8963736387111261,
"recall3": 0.9745424946671157,
"recall5": 0.9820366004266308,
"recall8": 0.9912709105198159,
"evidence_count_in_top3": 1.0851577411025037,
"hit3": 0.9761984955652857,
"mrr": 0.9670023550641874,
"ndcg3": 0.9642642772784563,
"mean_first_positive_rank": 1.1905243067250477,
"all_evidence_in_top3": 0.9728864937689458,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9705,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 200.33457717921095,
"median_score_margin": 219.4529800415039,
"negative_margin_rate": 0.04810822948242955,
"hop_confusion": {
"true_0_pred_0": 4106,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 15814,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999979790499,
"unknown_mean_need_prob": 3.8975788264237145e-06,
"thr0.30": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.40": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.60": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.70": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.80": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
}
},
"by_family": {
"benchmark_qa": {
"episodes": 64,
"answerable_episodes": 0,
"unknown_episodes": 64,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 64
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 3.895038815020929e-06,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 64,
"fp": 0,
"fn": 0
}
}
},
"mega_validation": {
"episodes": 20000,
"answerable_episodes": 16000,
"unknown_episodes": 4000,
"top1": 0.984375,
"recall1": 0.9225625,
"recall3": 0.99815625,
"recall5": 0.9995,
"recall8": 0.99959375,
"evidence_count_in_top3": 1.1213125,
"hit3": 1.0,
"mrr": 0.9921354166666666,
"ndcg3": 0.9914167841247224,
"mean_first_positive_rank": 1.0159375,
"all_evidence_in_top3": 0.9963125,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9705,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 217.4091286659348,
"median_score_margin": 228.10742282867432,
"negative_margin_rate": 0.015625,
"hop_confusion": {
"true_0_pred_0": 4000,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 14000,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999999627471,
"unknown_mean_need_prob": 6.477803048274694e-11,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 16000,
"tn": 4000,
"fp": 0,
"fn": 0
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.6754540128881078,
"recall1": 0.6754540128881078,
"recall3": 0.7592267135325131,
"recall5": 0.8189806678383128,
"recall8": 0.9127123608670181,
"evidence_count_in_top3": 0.7592267135325131,
"hit3": 0.7592267135325131,
"mrr": 0.7481470978152184,
"ndcg3": 0.7247037236923608,
"mean_first_positive_rank": 2.7809021675454013,
"all_evidence_in_top3": 0.7592267135325131,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 52.43232965865677,
"median_score_margin": 72.44926834106445,
"negative_margin_rate": 0.3245459871118922,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1707
},
"known_mean_need_prob": 0.9999999792588407,
"unknown_mean_need_prob": 0.0003750932634526786,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 1707,
"tn": 21,
"fp": 0,
"fn": 0
}
}
},
"native_memory": {
"episodes": 128,
"answerable_episodes": 107,
"unknown_episodes": 21,
"top1": 0.5046728971962616,
"recall1": 0.5046728971962616,
"recall3": 0.8785046728971962,
"recall5": 0.9719626168224299,
"recall8": 1.0,
"evidence_count_in_top3": 0.8785046728971962,
"hit3": 0.8785046728971962,
"mrr": 0.7002447708055184,
"ndcg3": 0.7258507766355217,
"mean_first_positive_rank": 1.925233644859813,
"all_evidence_in_top3": 0.8785046728971962,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 6.655256898882233,
"median_score_margin": 1.441606044769287,
"negative_margin_rate": 0.4953271028037383,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 107
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": 0.0003750932634526786,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 107,
"tn": 21,
"fp": 0,
"fn": 0
}
}
}
},
"by_category": {
"conflict_update": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9945,
"recall1": 0.9945,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99725,
"ndcg3": 0.9979701136446429,
"mean_first_positive_rank": 1.0055,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 243.0819873433709,
"median_score_margin": 235.45626258850098,
"negative_margin_rate": 0.0055,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_128": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9785,
"recall1": 0.9785,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.98925,
"ndcg3": 0.9920649897017864,
"mean_first_positive_rank": 1.0215,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 229.82924858646467,
"median_score_margin": 248.98048400878906,
"negative_margin_rate": 0.0215,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_32": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9845,
"recall1": 0.9845,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9921666666666666,
"ndcg3": 0.9942139463035718,
"mean_first_positive_rank": 1.016,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 232.02638172297924,
"median_score_margin": 249.2184543609619,
"negative_margin_rate": 0.0155,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"forget_correction": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 1.2955399044485864e-10,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
},
"long_context": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9815,
"recall1": 0.9815,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9906666666666666,
"ndcg3": 0.9931067355642863,
"mean_first_positive_rank": 1.019,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 211.99539810842276,
"median_score_margin": 225.90464401245117,
"negative_margin_rate": 0.0185,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"multi_hop": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.989,
"recall1": 0.4945,
"recall3": 0.98525,
"recall5": 0.996,
"recall8": 0.99675,
"evidence_count_in_top3": 1.9705,
"hit3": 1.0,
"mrr": 0.9945,
"ndcg3": 0.9737356054745618,
"mean_first_positive_rank": 1.011,
"all_evidence_in_top3": 0.9705,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9705,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 234.29880876231567,
"median_score_margin": 254.67290210723877,
"negative_margin_rate": 0.011,
"hop_confusion": {
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999997019767,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.997,
"recall1": 0.997,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9985,
"ndcg3": 0.9988927892607143,
"mean_first_positive_rank": 1.003,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 155.67033711057155,
"median_score_margin": 155.824444770813,
"negative_margin_rate": 0.003,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"random_position": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.981,
"recall1": 0.981,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9905,
"ndcg3": 0.9929876653178578,
"mean_first_positive_rank": 1.019,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 217.68418576198818,
"median_score_margin": 236.54215621948242,
"negative_margin_rate": 0.019,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"single_fact": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.969,
"recall1": 0.969,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.98425,
"ndcg3": 0.988362427730358,
"mean_first_positive_rank": 1.0325,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 214.68668193136529,
"median_score_margin": 221.31105995178223,
"negative_margin_rate": 0.031,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"unknown": {
"episodes": 1920,
"answerable_episodes": 1814,
"unknown_episodes": 106,
"top1": 0.665380374862183,
"recall1": 0.665380374862183,
"recall3": 0.7662624035281147,
"recall5": 0.8280044101433297,
"recall8": 0.9178610804851157,
"evidence_count_in_top3": 0.7662624035281147,
"hit3": 0.7662624035281147,
"mrr": 0.745321547104062,
"ndcg3": 0.7247713833753366,
"mean_first_positive_rank": 2.7304299889746417,
"all_evidence_in_top3": 0.7662624035281147,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 49.73213848704934,
"median_score_margin": 64.13029479980469,
"negative_margin_rate": 0.33461962513781696,
"hop_confusion": {
"true_0_pred_0": 106,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1814
},
"known_mean_need_prob": 0.9999999804822718,
"unknown_mean_need_prob": 0.0001509735806525834,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 1814,
"tn": 106,
"fp": 0,
"fn": 0
}
}
},
"unknown_abstention": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 2.0705206352511507e-15,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
}
}
},
"XL-512 v6 best": {
"checkpoint": "checkpoints\\router_v6_xl512\\router_best.pt",
"router": {
"kind": "router_xl",
"step": 36000,
"arch_config": {
"arch": "router_xl",
"arch_version": 1,
"hidden_size": 2560,
"router_dim": 512,
"num_heads": 8,
"max_hops": 3,
"encoder_layers": 2,
"encoder_hidden": 512,
"pair_blocks": 1,
"pair_hidden": 512,
"pair_expansion": 2,
"pair_dropout": 0.05,
"use_interaction": true,
"policy_layers": 2,
"policy_hidden": 512,
"policy_dropout": 0.0,
"learnable_cosine_scale": true
},
"parameters": 7898127,
"checkpoint_bytes": 99014505
},
"storage": {
"address_bytes_per_record": 2048,
"address_mb_per_1m_records": 2048,
"checkpoint_bytes": 99014505
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 1.5855000019655563,
"single_query_latency_ms_p50": 1.5855000019655563,
"single_query_latency_ms_p95": 1.6239000015048077,
"queries_per_second": 630.7158617220388,
"batched_qps_64": 39909.70428801726,
"batched_qps_256": 94647.25411335309,
"batched_latency_ms_64": 0.025056562503777968,
"samples": 5,
"source": "bench_router_latency.py (interleaved round-robin medians)"
},
"cpu": {
"single_query_latency_ms_mean": 1.7938500001037028,
"single_query_latency_ms_p50": 1.812500002415618,
"single_query_latency_ms_p95": 2.0920999995723832,
"queries_per_second": 557.4602112451932,
"samples": 30,
"batched_qps_64": 1675.9990132534788,
"batched_latency_ms_64": 0.5966590625007484,
"batched_qps_256": 1750.842524579076,
"batched_latency_ms_256": 0.5711535937479084
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.957842146626249,
"recall1": 0.9024082182553048,
"recall3": 0.9758897496351184,
"recall5": 0.9834399910183003,
"recall8": 0.9905411474121477,
"evidence_count_in_top3": 1.0808914337038285,
"hit3": 0.9831593128999663,
"mrr": 0.9724377931581535,
"ndcg3": 0.9664212737428025,
"mean_first_positive_rank": 1.1337150555742674,
"all_evidence_in_top3": 0.9686201863702706,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.8705,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 45.56406442258217,
"median_score_margin": 47.89992094039917,
"negative_margin_rate": 0.042157853373750984,
"hop_confusion": {
"true_0_pred_0": 4106,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 15814,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999977515257,
"unknown_mean_need_prob": 9.165001448203187e-06,
"thr0.30": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.40": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.60": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.70": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.80": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
}
},
"by_family": {
"benchmark_qa": {
"episodes": 64,
"answerable_episodes": 0,
"unknown_episodes": 64,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 64
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 1.130629900134017e-05,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 64,
"fp": 0,
"fn": 0
}
}
},
"mega_validation": {
"episodes": 20000,
"answerable_episodes": 16000,
"unknown_episodes": 4000,
"top1": 0.98375,
"recall1": 0.92203125,
"recall3": 0.99190625,
"recall5": 0.994875,
"recall8": 0.99640625,
"evidence_count_in_top3": 1.1088125,
"hit3": 1.0,
"mrr": 0.99184375,
"ndcg3": 0.9863265131728122,
"mean_first_positive_rank": 1.0164375,
"all_evidence_in_top3": 0.9838125,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.8705,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 48.28189839535998,
"median_score_margin": 49.04537320137024,
"negative_margin_rate": 0.01625,
"hop_confusion": {
"true_0_pred_0": 4000,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 14000,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999999552965,
"unknown_mean_need_prob": 2.862315262235959e-10,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 16000,
"tn": 4000,
"fp": 0,
"fn": 0
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.7457527826596367,
"recall1": 0.7457527826596367,
"recall3": 0.8347978910369068,
"recall5": 0.8763913298183948,
"recall8": 0.9355594610427651,
"evidence_count_in_top3": 0.8347978910369068,
"hit3": 0.8347978910369068,
"mrr": 0.8100220298155272,
"ndcg3": 0.7979455040170743,
"mean_first_positive_rank": 2.16813122437024,
"all_evidence_in_top3": 0.8347978910369068,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 22.8850553075215,
"median_score_margin": 27.899999618530273,
"negative_margin_rate": 0.2542472173403632,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1707
},
"known_mean_need_prob": 0.9999999769542673,
"unknown_mean_need_prob": 0.0008787320924793242,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 1707,
"tn": 21,
"fp": 0,
"fn": 0
}
}
},
"native_memory": {
"episodes": 128,
"answerable_episodes": 107,
"unknown_episodes": 21,
"top1": 0.4672897196261682,
"recall1": 0.4672897196261682,
"recall3": 0.8317757009345794,
"recall5": 0.9813084112149533,
"recall8": 0.9906542056074766,
"evidence_count_in_top3": 0.8317757009345794,
"hit3": 0.8317757009345794,
"mrr": 0.6616751628433871,
"ndcg3": 0.6776764890854534,
"mean_first_positive_rank": 2.1682242990654204,
"all_evidence_in_top3": 0.8317757009345794,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 0.9633634409336286,
"median_score_margin": -1.062392234802246,
"negative_margin_rate": 0.5327102803738317,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 107
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": 0.0008787320924793242,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 107,
"tn": 21,
"fp": 0,
"fn": 0
}
}
}
},
"by_category": {
"conflict_update": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.992,
"recall1": 0.992,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.996,
"ndcg3": 0.9970474380285717,
"mean_first_positive_rank": 1.008,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 67.69295654568822,
"median_score_margin": 69.3216142654419,
"negative_margin_rate": 0.008,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_128": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9795,
"recall1": 0.9795,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.98975,
"ndcg3": 0.9924340599482149,
"mean_first_positive_rank": 1.0205,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 39.98083441020921,
"median_score_margin": 44.252100229263306,
"negative_margin_rate": 0.0205,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_32": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9805,
"recall1": 0.9805,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99025,
"ndcg3": 0.9928031301946435,
"mean_first_positive_rank": 1.0195,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 40.70788796337694,
"median_score_margin": 44.594672441482544,
"negative_margin_rate": 0.0195,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"forget_correction": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 5.724454182644648e-10,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
},
"long_context": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9815,
"recall1": 0.9815,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9906666666666666,
"ndcg3": 0.9931067355642863,
"mean_first_positive_rank": 1.019,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 37.93500039816182,
"median_score_margin": 40.258564949035645,
"negative_margin_rate": 0.0185,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"multi_hop": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9875,
"recall1": 0.49375,
"recall3": 0.93525,
"recall5": 0.959,
"recall8": 0.97125,
"evidence_count_in_top3": 1.8705,
"hit3": 1.0,
"mrr": 0.9936666666666666,
"ndcg3": 0.9341087890914226,
"mean_first_positive_rank": 1.013,
"all_evidence_in_top3": 0.8705,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.8705,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 47.13378771021217,
"median_score_margin": 51.072168588638306,
"negative_margin_rate": 0.0125,
"hop_confusion": {
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9999999996423722,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9965,
"recall1": 0.9965,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99825,
"ndcg3": 0.9987082541375002,
"mean_first_positive_rank": 1.0035,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 61.5167154010348,
"median_score_margin": 62.94392251968384,
"negative_margin_rate": 0.0035,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"random_position": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9815,
"recall1": 0.9815,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99075,
"ndcg3": 0.9931722004410719,
"mean_first_positive_rank": 1.0185,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 38.85421975574363,
"median_score_margin": 43.87380111217499,
"negative_margin_rate": 0.0185,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"single_fact": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.971,
"recall1": 0.971,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9854166666666666,
"ndcg3": 0.9892314979767864,
"mean_first_positive_rank": 1.0295,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 52.43378497845307,
"median_score_margin": 59.86355948448181,
"negative_margin_rate": 0.029,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 1.0,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"unknown": {
"episodes": 1920,
"answerable_episodes": 1814,
"unknown_episodes": 106,
"top1": 0.7293274531422271,
"recall1": 0.7293274531422271,
"recall3": 0.834619625137817,
"recall5": 0.88257993384785,
"recall8": 0.9388092613009923,
"evidence_count_in_top3": 0.834619625137817,
"hit3": 0.834619625137817,
"mrr": 0.8012716909147449,
"ndcg3": 0.7908513559477891,
"mean_first_positive_rank": 2.1681367144432193,
"all_evidence_in_top3": 0.834619625137817,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 21.591989690253087,
"median_score_margin": 24.97784423828125,
"negative_margin_rate": 0.27067254685777287,
"hop_confusion": {
"true_0_pred_0": 106,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1814
},
"known_mean_need_prob": 0.9999999783136353,
"unknown_mean_need_prob": 0.0003550033115114848,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 1814,
"tn": 106,
"fp": 0,
"fn": 0
}
}
},
"unknown_abstention": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 1.7634182726834057e-14,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
}
}
},
"V2-512 v6 best": {
"checkpoint": "checkpoints\\router_v6_v2_512\\router_best.pt",
"router": {
"kind": "router_v2",
"step": 2000,
"arch_config": {
"arch": "router_v2",
"hidden_size": 2560,
"router_dim": 512,
"num_heads": 8,
"max_hops": 3
},
"parameters": 4741902,
"checkpoint_bytes": 56922573
},
"storage": {
"address_bytes_per_record": 2048,
"address_mb_per_1m_records": 2048,
"checkpoint_bytes": 56922573
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 1.162250000561471,
"single_query_latency_ms_p50": 1.162250000561471,
"single_query_latency_ms_p95": 1.1778999996749917,
"queries_per_second": 860.4000856243581,
"batched_qps_64": 53757.11860745415,
"batched_qps_256": 154085.05975293368,
"batched_latency_ms_64": 0.01860218750380227,
"samples": 5,
"source": "bench_router_latency.py (interleaved round-robin medians)"
},
"cpu": {
"single_query_latency_ms_mean": 0.8890533339581452,
"single_query_latency_ms_p50": 0.8698999990883749,
"single_query_latency_ms_p95": 1.0765000006358605,
"queries_per_second": 1124.7919127055181,
"samples": 30,
"batched_qps_64": 3131.6700902508155,
"batched_latency_ms_64": 0.31931843750498956,
"batched_qps_256": 3247.1683677121205,
"batched_latency_ms_256": 0.30796062499973686
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.950488379925901,
"recall1": 0.8971034018187942,
"recall3": 0.9884079937128102,
"recall5": 0.9934882676546536,
"recall8": 0.9961266419669923,
"evidence_count_in_top3": 1.0986302907825305,
"hit3": 0.9904569439766476,
"mrr": 0.9711571604054701,
"ndcg3": 0.9723689530227545,
"mean_first_positive_rank": 1.1156393847535646,
"all_evidence_in_top3": 0.9863590434489727,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9635,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 12.36926158705623,
"median_score_margin": 12.002254128456116,
"negative_margin_rate": 0.049511620074099025,
"hop_confusion": {
"true_0_pred_0": 4106,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 15814,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9988408471058552,
"unknown_mean_need_prob": 0.006478862109410833,
"thr0.30": {
"need_precision": 0.999326825984517,
"need_recall": 1.0,
"need_f1": 0.9996632996632996,
"specificity_unknown_refusal": 0.9970774476376035,
"abstention_accuracy": 0.9994525547445255,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0029225523623964927,
"counts": {
"tp": 17814,
"tn": 4094,
"fp": 12,
"fn": 0
}
},
"thr0.40": {
"need_precision": 0.999326825984517,
"need_recall": 1.0,
"need_f1": 0.9996632996632996,
"specificity_unknown_refusal": 0.9970774476376035,
"abstention_accuracy": 0.9994525547445255,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0029225523623964927,
"counts": {
"tp": 17814,
"tn": 4094,
"fp": 12,
"fn": 0
}
},
"thr0.50": {
"need_precision": 0.9995511165974638,
"need_recall": 1.0,
"need_f1": 0.999775507913346,
"specificity_unknown_refusal": 0.9980516317584024,
"abstention_accuracy": 0.9996350364963503,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.001948368241597662,
"counts": {
"tp": 17814,
"tn": 4098,
"fp": 8,
"fn": 0
}
},
"thr0.60": {
"need_precision": 0.9998877413560844,
"need_recall": 1.0,
"need_f1": 0.9999438675273646,
"specificity_unknown_refusal": 0.9995129079396006,
"abstention_accuracy": 0.9999087591240876,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0004870920603994155,
"counts": {
"tp": 17814,
"tn": 4104,
"fp": 2,
"fn": 0
}
},
"thr0.70": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
},
"thr0.80": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 17814,
"tn": 4106,
"fp": 0,
"fn": 0
}
}
},
"by_family": {
"benchmark_qa": {
"episodes": 64,
"answerable_episodes": 0,
"unknown_episodes": 64,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 64
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.0904978112084791,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 64,
"fp": 0,
"fn": 0
}
}
},
"mega_validation": {
"episodes": 20000,
"answerable_episodes": 16000,
"unknown_episodes": 4000,
"top1": 0.97,
"recall1": 0.9105625,
"recall3": 0.99771875,
"recall5": 0.9999375,
"recall8": 1.0,
"evidence_count_in_top3": 1.1204375,
"hit3": 1.0,
"mrr": 0.9849166666666667,
"ndcg3": 0.9857856193167982,
"mean_first_positive_rank": 1.0305,
"all_evidence_in_top3": 0.9954375,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9635,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 13.542529965203517,
"median_score_margin": 12.727105312049389,
"negative_margin_rate": 0.03,
"hop_confusion": {
"true_0_pred_0": 4000,
"true_0_pred_1": 0,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 14000,
"true_1_pred_2": 0,
"true_2_pred_0": 0,
"true_2_pred_1": 0,
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9990760059095919,
"unknown_mean_need_prob": 0.0025661538663334796,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 16000,
"tn": 4000,
"fp": 0,
"fn": 0
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.8125366139425894,
"recall1": 0.8125366139425894,
"recall3": 0.9244288224956063,
"recall5": 0.9502050380785003,
"recall8": 0.971294669009959,
"evidence_count_in_top3": 0.9244288224956063,
"hit3": 0.9244288224956063,
"mrr": 0.8742941912651278,
"ndcg3": 0.878837443896981,
"mean_first_positive_rank": 1.7176332747510252,
"all_evidence_in_top3": 0.9244288224956063,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 2.2332465364638656,
"median_score_margin": 2.495278835296631,
"negative_margin_rate": 0.18746338605741067,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1707
},
"known_mean_need_prob": 0.996912036259131,
"unknown_mean_need_prob": 0.2510888675848643,
"thr0.50": {
"need_precision": 0.9976621858562245,
"need_recall": 1.0,
"need_f1": 0.9988297249853716,
"specificity_unknown_refusal": 0.8095238095238095,
"abstention_accuracy": 0.9976851851851852,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.19047619047619047,
"counts": {
"tp": 1707,
"tn": 17,
"fp": 4,
"fn": 0
}
}
},
"native_memory": {
"episodes": 128,
"answerable_episodes": 107,
"unknown_episodes": 21,
"top1": 0.2336448598130841,
"recall1": 0.2336448598130841,
"recall3": 0.616822429906542,
"recall5": 0.719626168224299,
"recall8": 0.8130841121495327,
"evidence_count_in_top3": 0.616822429906542,
"hit3": 0.616822429906542,
"mrr": 0.4589420963252739,
"ndcg3": 0.4582719938918631,
"mean_first_positive_rank": 4.242990654205608,
"all_evidence_in_top3": 0.616822429906542,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -1.3701436372003823,
"median_score_margin": -0.6190404891967773,
"negative_margin_rate": 0.7663551401869159,
"hop_confusion": {
"true_0_pred_0": 21,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 107
},
"known_mean_need_prob": 0.9944477560364198,
"unknown_mean_need_prob": 0.2510888675848643,
"thr0.50": {
"need_precision": 0.963963963963964,
"need_recall": 1.0,
"need_f1": 0.9816513761467891,
"specificity_unknown_refusal": 0.8095238095238095,
"abstention_accuracy": 0.96875,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.19047619047619047,
"counts": {
"tp": 107,
"tn": 17,
"fp": 4,
"fn": 0
}
}
}
},
"by_category": {
"conflict_update": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9905,
"recall1": 0.9905,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99525,
"ndcg3": 0.9964938326589288,
"mean_first_positive_rank": 1.0095,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 16.784344019554556,
"median_score_margin": 14.332054734230042,
"negative_margin_rate": 0.0095,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9984212898612023,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_128": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.974,
"recall1": 0.974,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.987,
"ndcg3": 0.9904041735928579,
"mean_first_positive_rank": 1.026,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 14.32004562981613,
"median_score_margin": 12.83707270026207,
"negative_margin_rate": 0.026,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9998619016408921,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"distractor_32": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.969,
"recall1": 0.969,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9844166666666666,
"ndcg3": 0.9884933574839294,
"mean_first_positive_rank": 1.0315,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 14.515158921549562,
"median_score_margin": 12.90774929523468,
"negative_margin_rate": 0.031,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9998633873462677,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"forget_correction": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.004985218498797622,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
},
"long_context": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.973,
"recall1": 0.973,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9864166666666666,
"ndcg3": 0.9899696384696437,
"mean_first_positive_rank": 1.0275,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 11.82059130709432,
"median_score_margin": 12.446841090917587,
"negative_margin_rate": 0.027,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9998804149031639,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"multi_hop": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.951,
"recall1": 0.4755,
"recall3": 0.98175,
"recall5": 0.9995,
"recall8": 1.0,
"evidence_count_in_top3": 1.9635,
"hit3": 1.0,
"mrr": 0.9753333333333334,
"ndcg3": 0.9571701608629515,
"mean_first_positive_rank": 1.05,
"all_evidence_in_top3": 0.9635,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9635,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 10.0001081677787,
"median_score_margin": 11.238629460334778,
"negative_margin_rate": 0.049,
"hop_confusion": {
"true_2_pred_2": 2000
},
"known_mean_need_prob": 0.9978087923228741,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9755,
"recall1": 0.9755,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.98775,
"ndcg3": 0.9909577789625007,
"mean_first_positive_rank": 1.0245,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 14.530858453191817,
"median_score_margin": 12.788989305496216,
"negative_margin_rate": 0.0245,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9972889910042286,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"random_position": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9745,
"recall1": 0.9745,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9871666666666666,
"ndcg3": 0.9905232438392865,
"mean_first_positive_rank": 1.026,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 13.56638970718137,
"median_score_margin": 13.154640913009644,
"negative_margin_rate": 0.0255,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9995595035254955,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"single_fact": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.9525,
"recall1": 0.9525,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.976,
"ndcg3": 0.982272768664287,
"mean_first_positive_rank": 1.049,
"all_evidence_in_top3": 1.0,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 12.802743515461684,
"median_score_margin": 12.365995168685913,
"negative_margin_rate": 0.0475,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999237666726113,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": 1.0,
"need_recall": 1.0,
"need_f1": 1.0,
"specificity_unknown_refusal": null,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 2000,
"tn": 0,
"fp": 0,
"fn": 0
}
}
},
"unknown": {
"episodes": 1920,
"answerable_episodes": 1814,
"unknown_episodes": 106,
"top1": 0.7783902976846747,
"recall1": 0.7783902976846747,
"recall3": 0.906284454244763,
"recall5": 0.9366041896361632,
"recall8": 0.961962513781698,
"evidence_count_in_top3": 0.906284454244763,
"hit3": 0.906284454244763,
"mrr": 0.8497943708910569,
"ndcg3": 0.8540301102968997,
"mean_first_positive_rank": 1.866593164277839,
"all_evidence_in_top3": 0.906284454244763,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 2.0206981634858754,
"median_score_margin": 2.3058557510375977,
"negative_margin_rate": 0.22160970231532526,
"hop_confusion": {
"true_0_pred_0": 106,
"true_0_pred_1": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 1814
},
"known_mean_need_prob": 0.9967666790464352,
"unknown_mean_need_prob": 0.1541282297727072,
"thr0.50": {
"need_precision": 0.995609220636663,
"need_recall": 1.0,
"need_f1": 0.9977997799779977,
"specificity_unknown_refusal": 0.9245283018867925,
"abstention_accuracy": 0.9958333333333333,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.07547169811320754,
"counts": {
"tp": 1814,
"tn": 98,
"fp": 8,
"fn": 0
}
}
},
"unknown_abstention": {
"episodes": 2000,
"answerable_episodes": 0,
"unknown_episodes": 2000,
"top1": null,
"recall1": null,
"recall3": null,
"recall5": null,
"recall8": null,
"evidence_count_in_top3": null,
"hit3": null,
"mrr": null,
"ndcg3": null,
"mean_first_positive_rank": null,
"all_evidence_in_top3": null,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": null,
"hop_under_prediction": null,
"hop_over_prediction": null,
"mean_score_margin": null,
"median_score_margin": null,
"negative_margin_rate": null,
"hop_confusion": {
"true_0_pred_0": 2000
},
"known_mean_need_prob": null,
"unknown_mean_need_prob": 0.00014708923386933748,
"thr0.50": {
"need_precision": null,
"need_recall": null,
"need_f1": null,
"specificity_unknown_refusal": 1.0,
"abstention_accuracy": 1.0,
"known_question_refusal_rate": null,
"unknown_question_read_rate": 0.0,
"counts": {
"tp": 0,
"tn": 2000,
"fp": 0,
"fn": 0
}
}
}
}
}
}