Files
natural-memory-nm21/router_scorecard_v6.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

6124 lines
196 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": 0.8439246668907193,
"single_query_latency_ms_p50": 0.8266000004368834,
"single_query_latency_ms_p95": 1.1234000012336764,
"queries_per_second": 1184.9398876848934,
"samples": 150,
"batched_qps_64": 72916.19191648123,
"batched_latency_ms_64": 0.013714375006657065,
"batched_qps_256": 264812.97583671653,
"batched_latency_ms_256": 0.003776249999987158
},
"cpu": {
"single_query_latency_ms_mean": 0.3935313334174377,
"single_query_latency_ms_p50": 0.3775000004679896,
"single_query_latency_ms_p95": 0.5684999996447004,
"queries_per_second": 2541.093720075529,
"samples": 150,
"batched_qps_64": 11881.908679212764,
"batched_latency_ms_64": 0.08416156250632412,
"batched_qps_256": 13117.630300343435,
"batched_latency_ms_256": 0.07623328124850559
}
},
"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
}
}
}
}
},
"V2-512 v3 best": {
"checkpoint": "H:\\Memory\\dynamic_memory_lab\\checkpoints\\natural_memory_v2_router_512\\router_best.pt",
"router": {
"kind": "router_v2",
"step": 1000,
"arch_config": {
"arch": "router_v2",
"hidden_size": 2560,
"router_dim": 512,
"num_heads": 8,
"max_hops": 3
},
"parameters": 4741902,
"checkpoint_bytes": 56922381
},
"storage": {
"address_bytes_per_record": 2048,
"address_mb_per_1m_records": 2048,
"checkpoint_bytes": 56922381
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 0.8789160000742413,
"single_query_latency_ms_p50": 0.7994000006874558,
"single_query_latency_ms_p95": 1.517499997135019,
"queries_per_second": 1137.7651560735392,
"samples": 150,
"batched_qps_64": 69808.02789170416,
"batched_latency_ms_64": 0.014325000006465416,
"batched_qps_256": 167235.0043735693,
"batched_latency_ms_256": 0.005979609375117434
},
"cpu": {
"single_query_latency_ms_mean": 0.8516033333095644,
"single_query_latency_ms_p50": 0.8521000017935876,
"single_query_latency_ms_p95": 1.0607999975036364,
"queries_per_second": 1174.2556198244615,
"samples": 150,
"batched_qps_64": 3215.094870427158,
"batched_latency_ms_64": 0.3110328124989792,
"batched_qps_256": 3174.0694397097395,
"batched_latency_ms_256": 0.31505296875025124
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.807510946446615,
"recall1": 0.7698439429662064,
"recall3": 0.9501515661839003,
"recall5": 0.9730829684517796,
"recall8": 0.984506567867969,
"evidence_count_in_top3": 1.052542943752105,
"hit3": 0.9584035028629168,
"mrr": 0.8845870538857731,
"ndcg3": 0.8953026178367461,
"mean_first_positive_rank": 1.6810935219490288,
"all_evidence_in_top3": 0.9418996295048838,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.8385,
"hop_accuracy": 0.8158759124087591,
"hop_accuracy_answerable": 0.8877287526664421,
"hop_under_prediction": 0.11227124733355788,
"hop_over_prediction": 0.0,
"mean_score_margin": 9.638220129697089,
"median_score_margin": 7.411673903465271,
"negative_margin_rate": 0.19248905355338497,
"hop_confusion": {
"true_0_pred_0": 2070,
"true_0_pred_1": 2036,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 15814,
"true_1_pred_2": 0,
"true_2_pred_0": 2000,
"true_2_pred_1": 0,
"true_2_pred_2": 0
},
"known_mean_need_prob": 0.8877228112198915,
"unknown_mean_need_prob": 0.6232877522355414,
"thr0.30": {
"need_precision": 0.8045772861057546,
"need_recall": 0.8900303132367801,
"need_f1": 0.8451492537313433,
"specificity_unknown_refusal": 0.06210423770092548,
"abstention_accuracy": 0.7349452554744526,
"known_question_refusal_rate": 0.10996968676321994,
"unknown_question_read_rate": 0.9378957622990746,
"counts": {
"tp": 15855,
"tn": 255,
"fp": 3851,
"fn": 1959
}
},
"thr0.40": {
"need_precision": 0.885845516159749,
"need_recall": 0.8877848882901089,
"need_f1": 0.8868141419239074,
"specificity_unknown_refusal": 0.5036531904529956,
"abstention_accuracy": 0.8158302919708029,
"known_question_refusal_rate": 0.1122151117098911,
"unknown_question_read_rate": 0.4963468095470044,
"counts": {
"tp": 15815,
"tn": 2068,
"fp": 2038,
"fn": 1999
}
},
"thr0.50": {
"need_precision": 0.8859383753501401,
"need_recall": 0.8877287526664421,
"need_f1": 0.8868326603858232,
"specificity_unknown_refusal": 0.504140282513395,
"abstention_accuracy": 0.8158759124087591,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.49585971748660496,
"counts": {
"tp": 15814,
"tn": 2070,
"fp": 2036,
"fn": 2000
}
},
"thr0.60": {
"need_precision": 0.8868326603858232,
"need_recall": 0.8877287526664421,
"need_f1": 0.8872804802782921,
"specificity_unknown_refusal": 0.5085241110569898,
"abstention_accuracy": 0.8166970802919709,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.49147588894301025,
"counts": {
"tp": 15814,
"tn": 2088,
"fp": 2018,
"fn": 2000
}
},
"thr0.70": {
"need_precision": 0.8876290974405029,
"need_recall": 0.8877287526664421,
"need_f1": 0.8876789222565254,
"specificity_unknown_refusal": 0.5124208475401851,
"abstention_accuracy": 0.8174270072992701,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.4875791524598149,
"counts": {
"tp": 15814,
"tn": 2104,
"fp": 2002,
"fn": 2000
}
},
"thr0.80": {
"need_precision": 0.9022136010953903,
"need_recall": 0.8877287526664421,
"need_f1": 0.8949125686152454,
"specificity_unknown_refusal": 0.5825621042377009,
"abstention_accuracy": 0.8305656934306569,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.41743789576229906,
"counts": {
"tp": 15814,
"tn": 2392,
"fp": 1714,
"fn": 2000
}
}
},
"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": 0.04609098758373875,
"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": 0.8444375,
"recall1": 0.8025,
"recall3": 0.9736875,
"recall5": 0.98640625,
"recall8": 0.986875,
"evidence_count_in_top3": 1.0876875,
"hit3": 0.982875,
"mrr": 0.9118529559094745,
"ndcg3": 0.9247008759665758,
"mean_first_positive_rank": 1.5631875,
"all_evidence_in_top3": 0.9645,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.8385,
"hop_accuracy": 0.8,
"hop_accuracy_answerable": 0.875,
"hop_under_prediction": 0.125,
"hop_over_prediction": 0.0,
"mean_score_margin": 10.705225070305634,
"median_score_margin": 8.861685633659363,
"negative_margin_rate": 0.1555625,
"hop_confusion": {
"true_0_pred_0": 2000,
"true_0_pred_1": 2000,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 14000,
"true_1_pred_2": 0,
"true_2_pred_0": 2000,
"true_2_pred_1": 0,
"true_2_pred_2": 0
},
"known_mean_need_prob": 0.8765438592466526,
"unknown_mean_need_prob": 0.6330925791561604,
"thr0.50": {
"need_precision": 0.875,
"need_recall": 0.875,
"need_f1": 0.875,
"specificity_unknown_refusal": 0.5,
"abstention_accuracy": 0.8,
"known_question_refusal_rate": 0.125,
"unknown_question_read_rate": 0.5,
"counts": {
"tp": 14000,
"tn": 2000,
"fp": 2000,
"fn": 2000
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.4856473345049795,
"recall1": 0.4856473345049795,
"recall3": 0.7375512595196251,
"recall5": 0.8482718219097832,
"recall8": 0.961335676625659,
"evidence_count_in_top3": 0.7375512595196251,
"hit3": 0.7375512595196251,
"mrr": 0.6442412517298798,
"ndcg3": 0.6345330968470215,
"mean_first_positive_rank": 2.754540128881078,
"all_evidence_in_top3": 0.7375512595196251,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.9895833333333334,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 0.24317214623591182,
"median_score_margin": -0.0259857177734375,
"negative_margin_rate": 0.5143526654950205,
"hop_confusion": {
"true_0_pred_0": 3,
"true_0_pred_1": 18,
"true_1_pred_0": 0,
"true_1_pred_1": 1707
},
"known_mean_need_prob": 0.9860711831339773,
"unknown_mean_need_prob": 0.5690326392650604,
"thr0.50": {
"need_precision": 0.9895652173913043,
"need_recall": 1.0,
"need_f1": 0.9947552447552448,
"specificity_unknown_refusal": 0.14285714285714285,
"abstention_accuracy": 0.9895833333333334,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.8571428571428571,
"counts": {
"tp": 1707,
"tn": 3,
"fp": 18,
"fn": 0
}
}
},
"native_memory": {
"episodes": 128,
"answerable_episodes": 107,
"unknown_episodes": 21,
"top1": 0.4205607476635514,
"recall1": 0.4205607476635514,
"recall3": 0.822429906542056,
"recall5": 0.9719626168224299,
"recall8": 1.0,
"evidence_count_in_top3": 0.822429906542056,
"hit3": 0.822429906542056,
"mrr": 0.6417445482866044,
"ndcg3": 0.6594282463618242,
"mean_first_positive_rank": 2.1869158878504673,
"all_evidence_in_top3": 0.822429906542056,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.859375,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": -0.03217372047566922,
"median_score_margin": -0.08555316925048828,
"negative_margin_rate": 0.5794392523364486,
"hop_confusion": {
"true_0_pred_0": 3,
"true_0_pred_1": 18,
"true_1_pred_0": 0,
"true_1_pred_1": 107
},
"known_mean_need_prob": 0.9903635655608133,
"unknown_mean_need_prob": 0.5690326392650604,
"thr0.50": {
"need_precision": 0.856,
"need_recall": 1.0,
"need_f1": 0.9224137931034483,
"specificity_unknown_refusal": 0.14285714285714285,
"abstention_accuracy": 0.859375,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.8571428571428571,
"counts": {
"tp": 107,
"tn": 3,
"fp": 18,
"fn": 0
}
}
}
},
"by_category": {
"conflict_update": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.8605,
"recall1": 0.8605,
"recall3": 0.9995,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 0.9995,
"hit3": 0.9995,
"mrr": 0.9282083333333334,
"ndcg3": 0.9466935435803608,
"mean_first_positive_rank": 1.152,
"all_evidence_in_top3": 0.9995,
"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": 11.369050871308893,
"median_score_margin": 8.467366218566895,
"negative_margin_rate": 0.1395,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9782380353510379,
"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": 0.9215,
"recall1": 0.9215,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9599166666666666,
"ndcg3": 0.970373336887502,
"mean_first_positive_rank": 1.0835,
"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": 12.625836706157774,
"median_score_margin": 9.696466445922852,
"negative_margin_rate": 0.0785,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.959310285449028,
"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": 0.9215,
"recall1": 0.9215,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.95975,
"ndcg3": 0.9702424071339306,
"mean_first_positive_rank": 1.0845,
"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": 12.816123366106302,
"median_score_margin": 9.613193988800049,
"negative_margin_rate": 0.0785,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.959583455324173,
"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": 0.951540339499712,
"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": 0.782,
"recall1": 0.782,
"recall3": 0.989,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 0.989,
"hit3": 0.989,
"mrr": 0.8819583333333334,
"ndcg3": 0.9077580581071478,
"mean_first_positive_rank": 1.2795,
"all_evidence_in_top3": 0.989,
"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": 12.019205482423306,
"median_score_margin": 8.422776699066162,
"negative_margin_rate": 0.218,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9750196565389633,
"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": 0.671,
"recall1": 0.3355,
"recall3": 0.912,
"recall5": 0.99775,
"recall8": 1.0,
"evidence_count_in_top3": 1.824,
"hit3": 0.9855,
"mrr": 0.8213583333333333,
"ndcg3": 0.8246288558486577,
"mean_first_positive_rank": 1.4225,
"all_evidence_in_top3": 0.8385,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.8385,
"hop_accuracy": 0.0,
"hop_accuracy_answerable": 0.0,
"hop_under_prediction": 1.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 5.122980190102011,
"median_score_margin": 3.771735429763794,
"negative_margin_rate": 0.329,
"hop_confusion": {
"true_2_pred_0": 2000,
"true_2_pred_2": 0
},
"known_mean_need_prob": 0.1813941482268274,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": null,
"need_recall": 0.0,
"need_f1": null,
"specificity_unknown_refusal": null,
"abstention_accuracy": 0.0,
"known_question_refusal_rate": 1.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 0,
"tn": 0,
"fp": 0,
"fn": 2000
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.7725,
"recall1": 0.7725,
"recall3": 0.8895,
"recall5": 0.8935,
"recall8": 0.895,
"evidence_count_in_top3": 0.8895,
"hit3": 0.8895,
"mrr": 0.8329236472757954,
"ndcg3": 0.8441584402339315,
"mean_first_positive_rank": 4.2945,
"all_evidence_in_top3": 0.8895,
"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": 6.7703079060800375,
"median_score_margin": 4.96909373998642,
"negative_margin_rate": 0.2275,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9814486785233021,
"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": 0.942,
"recall1": 0.942,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.97025,
"ndcg3": 0.978004741816073,
"mean_first_positive_rank": 1.0625,
"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": 14.161037405148148,
"median_score_margin": 10.101282358169556,
"negative_margin_rate": 0.058,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9777550499141217,
"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": 0.8845,
"recall1": 0.8845,
"recall3": 0.9995,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 0.9995,
"hit3": 0.9995,
"mrr": 0.9404583333333334,
"ndcg3": 0.9557476241250031,
"mean_first_positive_rank": 1.1265,
"all_evidence_in_top3": 0.9995,
"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": 10.757258635118603,
"median_score_margin": 8.949375867843628,
"negative_margin_rate": 0.1155,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9996015646457672,
"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": 0.48180815876515987,
"recall1": 0.48180815876515987,
"recall3": 0.7425578831312017,
"recall5": 0.8555678059536935,
"recall8": 0.9636163175303197,
"evidence_count_in_top3": 0.7425578831312017,
"hit3": 0.7425578831312017,
"mrr": 0.6440939820118917,
"ndcg3": 0.6360015538470678,
"mean_first_positive_rank": 2.7210584343991178,
"all_evidence_in_top3": 0.7425578831312017,
"multi_positive_episodes": 0,
"all_evidence_in_top3_multi": null,
"hop_accuracy": 0.98125,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 0.2269306866228252,
"median_score_margin": -0.02995920181274414,
"negative_margin_rate": 0.5181918412348401,
"hop_confusion": {
"true_0_pred_0": 70,
"true_0_pred_1": 36,
"true_1_pred_0": 0,
"true_1_pred_1": 1814
},
"known_mean_need_prob": 0.98632437217459,
"unknown_mean_need_prob": 0.2532942835329417,
"thr0.50": {
"need_precision": 0.9805405405405405,
"need_recall": 1.0,
"need_f1": 0.990174672489083,
"specificity_unknown_refusal": 0.660377358490566,
"abstention_accuracy": 0.98125,
"known_question_refusal_rate": 0.0,
"unknown_question_read_rate": 0.33962264150943394,
"counts": {
"tp": 1814,
"tn": 70,
"fp": 36,
"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": 0.3146448188126087,
"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 v3 final": {
"checkpoint": "H:\\Memory\\dynamic_memory_lab\\checkpoints\\natural_memory_v2_router_512\\memory_router_v2.pt",
"router": {
"kind": "router_v2",
"step": null,
"arch_config": {
"arch": "router_v2",
"hidden_size": 2560,
"router_dim": 512,
"num_heads": 8,
"max_hops": 3
},
"parameters": 4741902,
"checkpoint_bytes": 18973433
},
"storage": {
"address_bytes_per_record": 2048,
"address_mb_per_1m_records": 2048,
"checkpoint_bytes": 18973433
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 0.9094866666661497,
"single_query_latency_ms_p50": 0.8472000008623581,
"single_query_latency_ms_p95": 1.27029999930528,
"queries_per_second": 1099.5213417098676,
"samples": 150,
"batched_qps_64": 67135.21454097527,
"batched_latency_ms_64": 0.014895312494900281,
"batched_qps_256": 168449.866383823,
"batched_latency_ms_256": 0.005936484376434237
},
"cpu": {
"single_query_latency_ms_mean": 0.8548113333866544,
"single_query_latency_ms_p50": 0.8628999967186246,
"single_query_latency_ms_p95": 1.0559000002103858,
"queries_per_second": 1169.8487852730339,
"samples": 150,
"batched_qps_64": 2969.669466452853,
"batched_latency_ms_64": 0.33673781250627144,
"batched_qps_256": 3091.357337730013,
"batched_latency_ms_256": 0.3234824999992725
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.7800044908498933,
"recall1": 0.7355731447176378,
"recall3": 0.928146401706523,
"recall5": 0.9529864151790727,
"recall8": 0.9671606601549343,
"evidence_count_in_top3": 1.0318288986190636,
"hit3": 0.936510609632873,
"mrr": 0.8612520819605932,
"ndcg3": 0.8691557053195522,
"mean_first_positive_rank": 1.9977545750533288,
"all_evidence_in_top3": 0.9197821937801729,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.849,
"hop_accuracy": 0.8175182481751825,
"hop_accuracy_answerable": 0.8877287526664421,
"hop_under_prediction": 0.11227124733355788,
"hop_over_prediction": 0.0,
"mean_score_margin": 53.160489893678886,
"median_score_margin": 48.923973083496094,
"negative_margin_rate": 0.21999550915010666,
"hop_confusion": {
"true_0_pred_0": 2106,
"true_0_pred_1": 2000,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 15814,
"true_1_pred_2": 0,
"true_2_pred_0": 2000,
"true_2_pred_1": 0,
"true_2_pred_2": 0
},
"known_mean_need_prob": 0.8877236850807078,
"unknown_mean_need_prob": 0.4528151513680766,
"thr0.30": {
"need_precision": 0.8877287526664421,
"need_recall": 0.8877287526664421,
"need_f1": 0.8877287526664421,
"specificity_unknown_refusal": 0.5129079396005846,
"abstention_accuracy": 0.8175182481751825,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.4870920603994155,
"counts": {
"tp": 15814,
"tn": 2106,
"fp": 2000,
"fn": 2000
}
},
"thr0.40": {
"need_precision": 0.8877287526664421,
"need_recall": 0.8877287526664421,
"need_f1": 0.8877287526664421,
"specificity_unknown_refusal": 0.5129079396005846,
"abstention_accuracy": 0.8175182481751825,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.4870920603994155,
"counts": {
"tp": 15814,
"tn": 2106,
"fp": 2000,
"fn": 2000
}
},
"thr0.50": {
"need_precision": 0.8877287526664421,
"need_recall": 0.8877287526664421,
"need_f1": 0.8877287526664421,
"specificity_unknown_refusal": 0.5129079396005846,
"abstention_accuracy": 0.8175182481751825,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.4870920603994155,
"counts": {
"tp": 15814,
"tn": 2106,
"fp": 2000,
"fn": 2000
}
},
"thr0.60": {
"need_precision": 0.9022136010953903,
"need_recall": 0.8877287526664421,
"need_f1": 0.8949125686152454,
"specificity_unknown_refusal": 0.5825621042377009,
"abstention_accuracy": 0.8305656934306569,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.41743789576229906,
"counts": {
"tp": 15814,
"tn": 2392,
"fp": 1714,
"fn": 2000
}
},
"thr0.70": {
"need_precision": 0.9022136010953903,
"need_recall": 0.8877287526664421,
"need_f1": 0.8949125686152454,
"specificity_unknown_refusal": 0.5825621042377009,
"abstention_accuracy": 0.8305656934306569,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.41743789576229906,
"counts": {
"tp": 15814,
"tn": 2392,
"fp": 1714,
"fn": 2000
}
},
"thr0.80": {
"need_precision": 0.9022136010953903,
"need_recall": 0.8877287526664421,
"need_f1": 0.8949125686152454,
"specificity_unknown_refusal": 0.5825621042377009,
"abstention_accuracy": 0.8305656934306569,
"known_question_refusal_rate": 0.11227124733355788,
"unknown_question_read_rate": 0.41743789576229906,
"counts": {
"tp": 15814,
"tn": 2392,
"fp": 1714,
"fn": 2000
}
}
},
"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": 8.474685865469522e-08,
"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": 0.8178125,
"recall1": 0.76834375,
"recall3": 0.9543125,
"recall5": 0.96865625,
"recall8": 0.9719375,
"evidence_count_in_top3": 1.06975,
"hit3": 0.963625,
"mrr": 0.8901514232628588,
"ndcg3": 0.9005039309554683,
"mean_first_positive_rank": 1.859375,
"all_evidence_in_top3": 0.945,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.849,
"hop_accuracy": 0.8,
"hop_accuracy_answerable": 0.875,
"hop_under_prediction": 0.125,
"hop_over_prediction": 0.0,
"mean_score_margin": 59.239713783336335,
"median_score_margin": 56.72071409225464,
"negative_margin_rate": 0.1821875,
"hop_confusion": {
"true_0_pred_0": 2000,
"true_0_pred_1": 2000,
"true_0_pred_2": 0,
"true_1_pred_0": 0,
"true_1_pred_1": 14000,
"true_1_pred_2": 0,
"true_2_pred_0": 2000,
"true_2_pred_1": 0,
"true_2_pred_2": 0
},
"known_mean_need_prob": 0.8749969797285945,
"unknown_mean_need_prob": 0.4644220104073552,
"thr0.50": {
"need_precision": 0.875,
"need_recall": 0.875,
"need_f1": 0.875,
"specificity_unknown_refusal": 0.5,
"abstention_accuracy": 0.8,
"known_question_refusal_rate": 0.125,
"unknown_question_read_rate": 0.5,
"counts": {
"tp": 14000,
"tn": 2000,
"fp": 2000,
"fn": 2000
}
}
},
"memory_policy": {
"episodes": 1728,
"answerable_episodes": 1707,
"unknown_episodes": 21,
"top1": 0.44639718804920914,
"recall1": 0.44639718804920914,
"recall3": 0.6942003514938488,
"recall5": 0.8096074985354423,
"recall8": 0.9209138840070299,
"evidence_count_in_top3": 0.6942003514938488,
"hit3": 0.6942003514938488,
"mrr": 0.6050143063165109,
"ndcg3": 0.5907781161122315,
"mean_first_positive_rank": 3.256590509666081,
"all_evidence_in_top3": 0.6942003514938488,
"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": -0.016779011952367244,
"median_score_margin": -1.933023452758789,
"negative_margin_rate": 0.5536028119507909,
"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": 0.9999754249386154,
"unknown_mean_need_prob": 0.03740391581197314,
"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": 0.4485981308411215,
"recall1": 0.4485981308411215,
"recall3": 0.7476635514018691,
"recall5": 0.897196261682243,
"recall8": 0.9906542056074766,
"evidence_count_in_top3": 0.7476635514018691,
"hit3": 0.7476635514018691,
"mrr": 0.6276859341867589,
"ndcg3": 0.6226036922563473,
"mean_first_positive_rank": 2.607476635514019,
"all_evidence_in_top3": 0.7476635514018691,
"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": -7.530951345635351,
"median_score_margin": -5.711280822753906,
"negative_margin_rate": 0.5514018691588785,
"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": 0.03740391581197314,
"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": 0.8025,
"recall1": 0.8025,
"recall3": 0.998,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 0.998,
"hit3": 0.998,
"mrr": 0.8965833333333334,
"ndcg3": 0.9225735229839335,
"mean_first_positive_rank": 1.2265,
"all_evidence_in_top3": 0.998,
"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": 60.32892245580256,
"median_score_margin": 50.02846574783325,
"negative_margin_rate": 0.1975,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999776878654957,
"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": 0.898,
"recall1": 0.898,
"recall3": 0.9995,
"recall5": 0.9995,
"recall8": 0.9995,
"evidence_count_in_top3": 0.9995,
"hit3": 0.9995,
"mrr": 0.9473787878787878,
"ndcg3": 0.9609264670821456,
"mean_first_positive_rank": 1.115,
"all_evidence_in_top3": 0.9995,
"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": 63.02626673298329,
"median_score_margin": 64.21059799194336,
"negative_margin_rate": 0.102,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999966541528702,
"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": 0.906,
"recall1": 0.906,
"recall3": 0.9995,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 0.9995,
"hit3": 0.9995,
"mrr": 0.9512916666666666,
"ndcg3": 0.9637480993000024,
"mean_first_positive_rank": 1.1045,
"all_evidence_in_top3": 0.9995,
"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": 63.26017077337205,
"median_score_margin": 64.34984970092773,
"negative_margin_rate": 0.094,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999967547059059,
"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": 0.9287911344170571,
"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": 0.6965,
"recall1": 0.6965,
"recall3": 0.8465,
"recall5": 0.8605,
"recall8": 0.8695,
"evidence_count_in_top3": 0.8465,
"hit3": 0.8465,
"mrr": 0.776365632860676,
"ndcg3": 0.7873425001821464,
"mean_first_positive_rank": 4.702,
"all_evidence_in_top3": 0.8465,
"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": 59.559096137337384,
"median_score_margin": 77.47207641601562,
"negative_margin_rate": 0.3035,
"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": 0.7915,
"recall1": 0.39575,
"recall3": 0.9235,
"recall5": 0.99575,
"recall8": 1.0,
"evidence_count_in_top3": 1.847,
"hit3": 0.998,
"mrr": 0.891,
"ndcg3": 0.8664546282365805,
"mean_first_positive_rank": 1.238,
"all_evidence_in_top3": 0.849,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.849,
"hop_accuracy": 0.0,
"hop_accuracy_answerable": 0.0,
"hop_under_prediction": 1.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 34.953358390808106,
"median_score_margin": 41.703819274902344,
"negative_margin_rate": 0.2085,
"hop_confusion": {
"true_2_pred_0": 2000,
"true_2_pred_2": 0
},
"known_mean_need_prob": 9.34347713071304e-06,
"unknown_mean_need_prob": null,
"thr0.50": {
"need_precision": null,
"need_recall": 0.0,
"need_f1": null,
"specificity_unknown_refusal": null,
"abstention_accuracy": 0.0,
"known_question_refusal_rate": 1.0,
"unknown_question_read_rate": null,
"counts": {
"tp": 0,
"tn": 0,
"fp": 0,
"fn": 2000
}
}
},
"paraphrase": {
"episodes": 2000,
"answerable_episodes": 2000,
"unknown_episodes": 0,
"top1": 0.625,
"recall1": 0.625,
"recall3": 0.8685,
"recall5": 0.8935,
"recall8": 0.9065,
"evidence_count_in_top3": 0.8685,
"hit3": 0.8685,
"mrr": 0.7501752986967395,
"ndcg3": 0.7711683990410768,
"mean_first_positive_rank": 3.293,
"all_evidence_in_top3": 0.8685,
"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": 22.522678687380626,
"median_score_margin": 15.791908264160156,
"negative_margin_rate": 0.375,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.999996542096138,
"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": 0.92,
"recall1": 0.92,
"recall3": 0.999,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 0.999,
"hit3": 0.999,
"mrr": 0.9584166666666666,
"ndcg3": 0.9687960125035735,
"mean_first_positive_rank": 1.09,
"all_evidence_in_top3": 0.999,
"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": 101.63241866916046,
"median_score_margin": 87.8995840549469,
"negative_margin_rate": 0.08,
"hop_confusion": {
"true_1_pred_1": 2000
},
"known_mean_need_prob": 0.9999988555312157,
"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": 0.903,
"recall1": 0.903,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.95,
"ndcg3": 0.9630218183142882,
"mean_first_positive_rank": 1.106,
"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": 68.63479841984622,
"median_score_margin": 62.0030552148819,
"negative_margin_rate": 0.097,
"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": 0.44652701212789414,
"recall1": 0.44652701212789414,
"recall3": 0.697353914002205,
"recall5": 0.814773980154355,
"recall8": 0.9250275633958104,
"evidence_count_in_top3": 0.697353914002205,
"hit3": 0.697353914002205,
"mrr": 0.606351607409188,
"ndcg3": 0.5926553689498392,
"mean_first_positive_rank": 3.2183020948180814,
"all_evidence_in_top3": 0.697353914002205,
"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": -0.46000747926442864,
"median_score_margin": -2.0159082412719727,
"negative_margin_rate": 0.5534729878721059,
"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": 0.9999768745150035,
"unknown_mean_need_prob": 0.01482047064058326,
"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": 5.2886397653310265e-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": 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": 0.8750353333986519,
"single_query_latency_ms_p50": 0.8467999978165608,
"single_query_latency_ms_p95": 1.2185999985376839,
"queries_per_second": 1142.8109949754637,
"samples": 150,
"batched_qps_64": 73617.37369428946,
"batched_latency_ms_64": 0.013583750001089356,
"batched_qps_256": 161941.27100652116,
"batched_latency_ms_256": 0.006175078124215361
},
"cpu": {
"single_query_latency_ms_mean": 0.921385333276703,
"single_query_latency_ms_p50": 0.8951000017987099,
"single_query_latency_ms_p95": 1.1334999981045257,
"queries_per_second": 1085.3222467126986,
"samples": 150,
"batched_qps_64": 2924.674104534912,
"batched_latency_ms_64": 0.3419184374934048,
"batched_qps_256": 3073.5827721960004,
"batched_latency_ms_256": 0.32535320312376825
}
},
"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
}
}
}
}
},
"V2-512 v6 final": {
"checkpoint": "checkpoints\\router_v6_v2_512\\memory_router_v2.pt",
"router": {
"kind": "router_v2",
"step": null,
"arch_config": {
"arch": "router_v2",
"hidden_size": 2560,
"router_dim": 512,
"num_heads": 8,
"max_hops": 3
},
"parameters": 4741902,
"checkpoint_bytes": 18973433
},
"storage": {
"address_bytes_per_record": 2048,
"address_mb_per_1m_records": 2048,
"checkpoint_bytes": 18973433
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 0.8725079999809774,
"single_query_latency_ms_p50": 0.850600001285784,
"single_query_latency_ms_p95": 1.0982999992847908,
"queries_per_second": 1146.121296334019,
"samples": 150,
"batched_qps_64": 61636.8434007473,
"batched_latency_ms_64": 0.016224062505898473,
"batched_qps_256": 164167.81025689258,
"batched_latency_ms_256": 0.006091328125990003
},
"cpu": {
"single_query_latency_ms_mean": 0.8932059997459874,
"single_query_latency_ms_p50": 0.876300000527408,
"single_query_latency_ms_p95": 1.1406000012357254,
"queries_per_second": 1119.5625648331775,
"samples": 150,
"batched_qps_64": 2943.6168587607826,
"batched_latency_ms_64": 0.3397181250079484,
"batched_qps_256": 3143.215455374193,
"batched_latency_ms_256": 0.3181455468762806
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.941843493881217,
"recall1": 0.8864657011339396,
"recall3": 0.9666273717300999,
"recall5": 0.9764230380599529,
"recall8": 0.984506567867969,
"evidence_count_in_top3": 1.07668126192882,
"hit3": 0.9688447288649377,
"mrr": 0.958724954839147,
"ndcg3": 0.9548399733910118,
"mean_first_positive_rank": 1.2825867295385651,
"all_evidence_in_top3": 0.9644100145952621,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9605,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 73.24820388849786,
"median_score_margin": 76.59281730651855,
"negative_margin_rate": 0.05815650611878298,
"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": 4.522551017238816e-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": 4.432710963581804e-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": 0.98375,
"recall1": 0.92209375,
"recall3": 0.99753125,
"recall5": 0.9999375,
"recall8": 1.0,
"evidence_count_in_top3": 1.1200625,
"hit3": 1.0,
"mrr": 0.99184375,
"ndcg3": 0.9902417180168516,
"mean_first_positive_rank": 1.0164375,
"all_evidence_in_top3": 0.9950625,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9605,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 81.85253369444608,
"median_score_margin": 83.23931503295898,
"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": 0.9999999999701977,
"unknown_mean_need_prob": 6.058075942262249e-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": 0.5770357352079672,
"recall1": 0.5770357352079672,
"recall3": 0.6859988283538372,
"recall5": 0.7562975981253661,
"recall8": 0.8388986526069128,
"evidence_count_in_top3": 0.6859988283538372,
"hit3": 0.6859988283538372,
"mrr": 0.6663638286240626,
"ndcg3": 0.6401845706816489,
"mean_first_positive_rank": 3.719976567076743,
"all_evidence_in_top3": 0.6859988283538372,
"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": -2.648404171182946,
"median_score_margin": 5.544002532958984,
"negative_margin_rate": 0.4229642647920328,
"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": 0.00043537282504942034,
"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": 0.4953271028037383,
"recall1": 0.4953271028037383,
"recall3": 0.822429906542056,
"recall5": 0.9719626168224299,
"recall8": 0.9906542056074766,
"evidence_count_in_top3": 0.822429906542056,
"hit3": 0.822429906542056,
"mrr": 0.6704980377877574,
"ndcg3": 0.680904070694264,
"mean_first_positive_rank": 2.196261682242991,
"all_evidence_in_top3": 0.822429906542056,
"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": -2.581393656329574,
"median_score_margin": -0.056890010833740234,
"negative_margin_rate": 0.5046728971962616,
"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": 0.00043537282504942034,
"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": 0.9935,
"recall1": 0.9935,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99675,
"ndcg3": 0.9976010433982145,
"mean_first_positive_rank": 1.0065,
"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": 64.64610423742235,
"median_score_margin": 60.52951920032501,
"negative_margin_rate": 0.0065,
"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": 0.98,
"recall1": 0.98,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99,
"ndcg3": 0.9926185950714291,
"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": 88.94486605204642,
"median_score_margin": 95.25923618674278,
"negative_margin_rate": 0.02,
"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": 0.9901666666666666,
"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": 90.02319332621246,
"median_score_margin": 96.11362898349762,
"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": 1.2115745555928435e-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": 0.977,
"recall1": 0.977,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9884166666666666,
"ndcg3": 0.9914459194553578,
"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": 78.39938361007721,
"median_score_margin": 81.58803009986877,
"negative_margin_rate": 0.023,
"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": 0.9865,
"recall1": 0.49325,
"recall3": 0.98025,
"recall5": 0.9995,
"recall8": 1.0,
"evidence_count_in_top3": 1.9605,
"hit3": 1.0,
"mrr": 0.99325,
"ndcg3": 0.9651268224740949,
"mean_first_positive_rank": 1.0135,
"all_evidence_in_top3": 0.9605,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.9605,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 104.36254599650204,
"median_score_margin": 118.10550880432129,
"negative_margin_rate": 0.0135,
"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": 0.9955,
"recall1": 0.9955,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99775,
"ndcg3": 0.9983391838910716,
"mean_first_positive_rank": 1.0045,
"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": 57.61693591082841,
"median_score_margin": 57.403422355651855,
"negative_margin_rate": 0.0045,
"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": 0.9825,
"recall1": 0.9825,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99125,
"ndcg3": 0.9935412706875005,
"mean_first_positive_rank": 1.0175,
"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": 84.6615379076004,
"median_score_margin": 90.02162456512451,
"negative_margin_rate": 0.0175,
"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": 86.1657025148794,
"median_score_margin": 91.7922420501709,
"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": 0.5722160970231532,
"recall1": 0.5722160970231532,
"recall3": 0.6940463065049615,
"recall5": 0.7690187431091511,
"recall8": 0.8478500551267916,
"evidence_count_in_top3": 0.6940463065049615,
"hit3": 0.6940463065049615,
"mrr": 0.666607687709242,
"ndcg3": 0.6425864375511913,
"mean_first_positive_rank": 3.630099228224917,
"all_evidence_in_top3": 0.6940463065049615,
"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": -2.644451511266016,
"median_score_margin": 4.475192725658417,
"negative_margin_rate": 0.42778390297684676,
"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": 0.00017518256748815932,
"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": 4.063285960620704e-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": 1.2161159999474573,
"single_query_latency_ms_p50": 1.1961999989580363,
"single_query_latency_ms_p95": 1.5090000015334226,
"queries_per_second": 822.2899789520122,
"samples": 150,
"batched_qps_64": 57386.75086015149,
"batched_latency_ms_64": 0.017425624992029043,
"batched_qps_256": 96746.88596045674,
"batched_latency_ms_256": 0.010336249999909342
},
"cpu": {
"single_query_latency_ms_mean": 1.6496533333702246,
"single_query_latency_ms_p50": 1.6746999972383492,
"single_query_latency_ms_p95": 2.007999999477761,
"queries_per_second": 606.1879667511782,
"samples": 150,
"batched_qps_64": 1660.2789164532574,
"batched_latency_ms_64": 0.6023084375101462,
"batched_qps_256": 1699.1487264831305,
"batched_latency_ms_256": 0.5885300000016969
}
},
"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
}
}
}
}
},
"XL-512 v6 final": {
"checkpoint": "checkpoints\\router_v6_xl512\\memory_router_xl.pt",
"router": {
"kind": "router_xl",
"step": null,
"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": 35802155
},
"storage": {
"address_bytes_per_record": 2048,
"address_mb_per_1m_records": 2048,
"checkpoint_bytes": 35802155
},
"latency": {
"cuda": {
"single_query_latency_ms_mean": 1.2505186665900208,
"single_query_latency_ms_p50": 1.2239999996381812,
"single_query_latency_ms_p95": 1.6236000010394491,
"queries_per_second": 799.6681910609555,
"samples": 150,
"batched_qps_64": 53291.58828255585,
"batched_latency_ms_64": 0.018764687490602228,
"batched_qps_256": 94345.88082639035,
"batched_latency_ms_256": 0.010599296876989683
},
"cpu": {
"single_query_latency_ms_mean": 1.717839333335481,
"single_query_latency_ms_p50": 1.7062999977497384,
"single_query_latency_ms_p95": 2.0226999986334704,
"queries_per_second": 582.1266171955253,
"samples": 150,
"batched_qps_64": 1690.8261059433491,
"batched_latency_ms_64": 0.5914268749961593,
"batched_qps_256": 1707.8817676137278,
"batched_latency_ms_256": 0.585520625000413
}
},
"metrics": {
"episodes": 21920,
"answerable_episodes": 17814,
"unknown_episodes": 4106,
"top1": 0.9502077018075671,
"recall1": 0.8948299090602897,
"recall3": 0.965336252385764,
"recall5": 0.9760862243179522,
"recall8": 0.9854608734703043,
"evidence_count_in_top3": 1.0657348153137982,
"hit3": 0.9772089367912877,
"mrr": 0.9664213223997663,
"ndcg3": 0.9560053718886551,
"mean_first_positive_rank": 1.1716066015493432,
"all_evidence_in_top3": 0.9534635679802402,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.7885,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 215.75331382084113,
"median_score_margin": 239.68688774108887,
"negative_margin_rate": 0.049792298192432916,
"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": 4.1244088529706584e-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": 4.264552800687521e-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": 0.984,
"recall1": 0.92234375,
"recall3": 0.98678125,
"recall5": 0.990875,
"recall8": 0.9933125,
"evidence_count_in_top3": 1.0985625,
"hit3": 1.0,
"mrr": 0.99196875,
"ndcg3": 0.9823026449115756,
"mean_first_positive_rank": 1.0161875,
"all_evidence_in_top3": 0.9735625,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.7885,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 235.8910635419197,
"median_score_margin": 250.7837634086609,
"negative_margin_rate": 0.016,
"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": 7.011441629130955e-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": 0.6690099589923842,
"recall1": 0.6690099589923842,
"recall3": 0.7709431751611013,
"recall5": 0.8388986526069128,
"recall8": 0.9121265377855887,
"evidence_count_in_top3": 0.7709431751611013,
"hit3": 0.7709431751611013,
"mrr": 0.7488416468644941,
"ndcg3": 0.7281836459473614,
"mean_first_positive_rank": 2.558289396602226,
"all_evidence_in_top3": 0.7709431751611013,
"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": 40.85260479970152,
"median_score_margin": 62.14678978919983,
"negative_margin_rate": 0.3309900410076157,
"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": 0.00039670502174734185,
"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": 0.38317757009345793,
"recall1": 0.38317757009345793,
"recall3": 0.8598130841121495,
"recall5": 0.9532710280373832,
"recall8": 0.9813084112149533,
"evidence_count_in_top3": 0.8598130841121495,
"hit3": 0.8598130841121495,
"mrr": 0.617352766651832,
"ndcg3": 0.6582046038050816,
"mean_first_positive_rank": 2.289719626168224,
"all_evidence_in_top3": 0.8598130841121495,
"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": -5.260566909736562,
"median_score_margin": -7.525552272796631,
"negative_margin_rate": 0.616822429906542,
"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": 0.00039670502174734185,
"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": 0.9925,
"recall1": 0.9925,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99625,
"ndcg3": 0.9972319731517859,
"mean_first_positive_rank": 1.0075,
"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": 272.5490972074904,
"median_score_margin": 288.63626432418823,
"negative_margin_rate": 0.0075,
"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": 0.98,
"recall1": 0.98,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.99,
"ndcg3": 0.9926185950714291,
"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": 215.3791910160221,
"median_score_margin": 239.17540168762207,
"negative_margin_rate": 0.02,
"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": 0.982,
"recall1": 0.982,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9909166666666667,
"ndcg3": 0.9932912706875006,
"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": 219.1182684260085,
"median_score_margin": 239.48388767242432,
"negative_margin_rate": 0.018,
"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": 1.4022628460683615e-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": 0.983,
"recall1": 0.983,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9914166666666666,
"ndcg3": 0.9936603409339291,
"mean_first_positive_rank": 1.0175,
"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": 211.44927242165804,
"median_score_margin": 233.2896122932434,
"negative_margin_rate": 0.017,
"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": 0.9865,
"recall1": 0.49325,
"recall3": 0.89425,
"recall5": 0.927,
"recall8": 0.9465,
"evidence_count_in_top3": 1.7885,
"hit3": 1.0,
"mrr": 0.99325,
"ndcg3": 0.9008760971390303,
"mean_first_positive_rank": 1.0135,
"all_evidence_in_top3": 0.7885,
"multi_positive_episodes": 2000,
"all_evidence_in_top3_multi": 0.7885,
"hop_accuracy": 1.0,
"hop_accuracy_answerable": 1.0,
"hop_under_prediction": 0.0,
"hop_over_prediction": 0.0,
"mean_score_margin": 274.43398497897016,
"median_score_margin": 301.56740951538086,
"negative_margin_rate": 0.0135,
"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": 0.995,
"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": 223.85322645658067,
"median_score_margin": 222.2765736579895,
"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": 0.99025,
"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": 225.64811652239038,
"median_score_margin": 241.96742022037506,
"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": 0.9725,
"recall1": 0.9725,
"recall3": 1.0,
"recall5": 1.0,
"recall8": 1.0,
"evidence_count_in_top3": 1.0,
"hit3": 1.0,
"mrr": 0.9861666666666666,
"ndcg3": 0.9897851033464293,
"mean_first_positive_rank": 1.028,
"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": 244.6973513062373,
"median_score_margin": 263.6290616989136,
"negative_margin_rate": 0.0275,
"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": 0.6521499448732084,
"recall1": 0.6521499448732084,
"recall3": 0.7761852260198456,
"recall5": 0.8456449834619625,
"recall8": 0.9162072767364939,
"evidence_count_in_top3": 0.7761852260198456,
"hit3": 0.7761852260198456,
"mrr": 0.7410856875575729,
"ndcg3": 0.7240558854681862,
"mean_first_positive_rank": 2.54244762954796,
"all_evidence_in_top3": 0.7761852260198456,
"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": 38.1325886073587,
"median_score_margin": 54.42159175872803,
"negative_margin_rate": 0.3478500551267916,
"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": 0.0001597598329493619,
"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": 2.5479757829358922e-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
}
}
}
}
}
}