- 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
1740 lines
54 KiB
JSON
1740 lines
54 KiB
JSON
{
|
|
"V2-128 v6 final": {
|
|
"checkpoint": "checkpoints\\router_v6_v2_128\\memory_router_v2.pt",
|
|
"router": {
|
|
"kind": "router_v2",
|
|
"step": null,
|
|
"arch_config": {
|
|
"arch": "router_v2",
|
|
"hidden_size": 2560,
|
|
"router_dim": 128,
|
|
"num_heads": 8,
|
|
"max_hops": 3
|
|
},
|
|
"parameters": 2037774,
|
|
"checkpoint_bytes": 8156921
|
|
},
|
|
"storage": {
|
|
"address_bytes_per_record": 512,
|
|
"address_mb_per_1m_records": 512,
|
|
"checkpoint_bytes": 8156921
|
|
},
|
|
"latency": {
|
|
"cuda": {
|
|
"single_query_latency_ms_mean": 0.8988993330664622,
|
|
"single_query_latency_ms_p50": 0.869700001203455,
|
|
"single_query_latency_ms_p95": 1.2395000012475066,
|
|
"queries_per_second": 1112.4716230333022,
|
|
"samples": 150,
|
|
"batched_qps_64": 67531.91938657848,
|
|
"batched_latency_ms_64": 0.014807812499384454,
|
|
"batched_qps_256": 237234.73279415478,
|
|
"batched_latency_ms_256": 0.004215234372395571
|
|
},
|
|
"cpu": {
|
|
"single_query_latency_ms_mean": 0.36343133309856057,
|
|
"single_query_latency_ms_p50": 0.3530000030878,
|
|
"single_query_latency_ms_p95": 0.5296000017551705,
|
|
"queries_per_second": 2751.5514181844237,
|
|
"samples": 150,
|
|
"batched_qps_64": 14346.173160954218,
|
|
"batched_latency_ms_64": 0.0697049999871524,
|
|
"batched_qps_256": 13080.658610801669,
|
|
"batched_latency_ms_256": 0.0764487500021005
|
|
}
|
|
},
|
|
"metrics": {
|
|
"episodes": 21920,
|
|
"answerable_episodes": 17814,
|
|
"unknown_episodes": 4106,
|
|
"top1": 0.9436959694622207,
|
|
"recall1": 0.8882059054676097,
|
|
"recall3": 0.9639609296059279,
|
|
"recall5": 0.973981138430448,
|
|
"recall8": 0.9864151790726395,
|
|
"evidence_count_in_top3": 1.0717974626698104,
|
|
"hit3": 0.9683956438756035,
|
|
"mrr": 0.9597054333626427,
|
|
"ndcg3": 0.9530820526370456,
|
|
"mean_first_positive_rank": 1.2573818345121814,
|
|
"all_evidence_in_top3": 0.9595262153362524,
|
|
"multi_positive_episodes": 2000,
|
|
"all_evidence_in_top3_multi": 0.921,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 51.315328869764656,
|
|
"median_score_margin": 56.234654903411865,
|
|
"negative_margin_rate": 0.05630403053777928,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 4106,
|
|
"true_0_pred_1": 0,
|
|
"true_0_pred_2": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 15814,
|
|
"true_1_pred_2": 0,
|
|
"true_2_pred_0": 0,
|
|
"true_2_pred_1": 0,
|
|
"true_2_pred_2": 2000
|
|
},
|
|
"known_mean_need_prob": 0.9999999984274064,
|
|
"unknown_mean_need_prob": 4.836221294362452e-06,
|
|
"thr0.30": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.40": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.60": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.70": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.80": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"by_family": {
|
|
"benchmark_qa": {
|
|
"episodes": 64,
|
|
"answerable_episodes": 0,
|
|
"unknown_episodes": 64,
|
|
"top1": null,
|
|
"recall1": null,
|
|
"recall3": null,
|
|
"recall5": null,
|
|
"recall8": null,
|
|
"evidence_count_in_top3": null,
|
|
"hit3": null,
|
|
"mrr": null,
|
|
"ndcg3": null,
|
|
"mean_first_positive_rank": null,
|
|
"all_evidence_in_top3": null,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": null,
|
|
"hop_under_prediction": null,
|
|
"hop_over_prediction": null,
|
|
"mean_score_margin": null,
|
|
"median_score_margin": null,
|
|
"negative_margin_rate": null,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 64
|
|
},
|
|
"known_mean_need_prob": null,
|
|
"unknown_mean_need_prob": 4.8406160974828055e-06,
|
|
"thr0.50": {
|
|
"need_precision": null,
|
|
"need_recall": null,
|
|
"need_f1": null,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": null,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 0,
|
|
"tn": 64,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"mega_validation": {
|
|
"episodes": 20000,
|
|
"answerable_episodes": 16000,
|
|
"unknown_episodes": 4000,
|
|
"top1": 0.9838125,
|
|
"recall1": 0.92203125,
|
|
"recall3": 0.9950625,
|
|
"recall5": 0.99978125,
|
|
"recall8": 0.999875,
|
|
"evidence_count_in_top3": 1.115125,
|
|
"hit3": 1.0,
|
|
"mrr": 0.991875,
|
|
"ndcg3": 0.9876817889701721,
|
|
"mean_first_positive_rank": 1.016375,
|
|
"all_evidence_in_top3": 0.990125,
|
|
"multi_positive_episodes": 2000,
|
|
"all_evidence_in_top3_multi": 0.921,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 57.110110210259094,
|
|
"median_score_margin": 59.90444374084473,
|
|
"negative_margin_rate": 0.0161875,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 4000,
|
|
"true_0_pred_1": 0,
|
|
"true_0_pred_2": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 14000,
|
|
"true_1_pred_2": 0,
|
|
"true_2_pred_0": 0,
|
|
"true_2_pred_1": 0,
|
|
"true_2_pred_2": 2000
|
|
},
|
|
"known_mean_need_prob": 0.9999999999701977,
|
|
"unknown_mean_need_prob": 6.379481702447893e-11,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 16000,
|
|
"tn": 4000,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"memory_policy": {
|
|
"episodes": 1728,
|
|
"answerable_episodes": 1707,
|
|
"unknown_episodes": 21,
|
|
"top1": 0.5940246045694201,
|
|
"recall1": 0.5940246045694201,
|
|
"recall3": 0.6842413591095489,
|
|
"recall5": 0.7352079671939075,
|
|
"recall8": 0.8611599297012302,
|
|
"evidence_count_in_top3": 0.6842413591095489,
|
|
"hit3": 0.6842413591095489,
|
|
"mrr": 0.6760273585431921,
|
|
"ndcg3": 0.6466497456649108,
|
|
"mean_first_positive_rank": 3.4463971880492092,
|
|
"all_evidence_in_top3": 0.6842413591095489,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 0.27154229221702947,
|
|
"median_score_margin": 7.975786209106445,
|
|
"negative_margin_rate": 0.40597539543058,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 21,
|
|
"true_0_pred_1": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 1707
|
|
},
|
|
"known_mean_need_prob": 0.9999999838679872,
|
|
"unknown_mean_need_prob": 0.0004654159529796483,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 1707,
|
|
"tn": 21,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"native_memory": {
|
|
"episodes": 128,
|
|
"answerable_episodes": 107,
|
|
"unknown_episodes": 21,
|
|
"top1": 0.5233644859813084,
|
|
"recall1": 0.5233644859813084,
|
|
"recall3": 0.7757009345794392,
|
|
"recall5": 0.9252336448598131,
|
|
"recall8": 0.9719626168224299,
|
|
"evidence_count_in_top3": 0.7757009345794392,
|
|
"hit3": 0.7757009345794392,
|
|
"mrr": 0.674896157840083,
|
|
"ndcg3": 0.6678873486315127,
|
|
"mean_first_positive_rank": 2.3738317757009346,
|
|
"all_evidence_in_top3": 0.7757009345794392,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": -0.8786688875929217,
|
|
"median_score_margin": 0.4756507873535156,
|
|
"negative_margin_rate": 0.4766355140186916,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 21,
|
|
"true_0_pred_1": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 107
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": 0.0004654159529796483,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 107,
|
|
"tn": 21,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
}
|
|
},
|
|
"by_category": {
|
|
"conflict_update": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.993,
|
|
"recall1": 0.993,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9965,
|
|
"ndcg3": 0.9974165082750002,
|
|
"mean_first_positive_rank": 1.007,
|
|
"all_evidence_in_top3": 1.0,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 54.4210530708991,
|
|
"median_score_margin": 56.92102670669556,
|
|
"negative_margin_rate": 0.007,
|
|
"hop_confusion": {
|
|
"true_1_pred_1": 2000
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": null,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": null,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": null,
|
|
"counts": {
|
|
"tp": 2000,
|
|
"tn": 0,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"distractor_128": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.975,
|
|
"recall1": 0.975,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9875,
|
|
"ndcg3": 0.9907732438392864,
|
|
"mean_first_positive_rank": 1.025,
|
|
"all_evidence_in_top3": 1.0,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 64.29146421498805,
|
|
"median_score_margin": 73.40720510482788,
|
|
"negative_margin_rate": 0.025,
|
|
"hop_confusion": {
|
|
"true_1_pred_1": 2000
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": null,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": null,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": null,
|
|
"counts": {
|
|
"tp": 2000,
|
|
"tn": 0,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"distractor_32": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.9845,
|
|
"recall1": 0.9845,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.99225,
|
|
"ndcg3": 0.9942794111803576,
|
|
"mean_first_positive_rank": 1.0155,
|
|
"all_evidence_in_top3": 1.0,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 65.52974343589041,
|
|
"median_score_margin": 73.97303533554077,
|
|
"negative_margin_rate": 0.0155,
|
|
"hop_confusion": {
|
|
"true_1_pred_1": 2000
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": null,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": null,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": null,
|
|
"counts": {
|
|
"tp": 2000,
|
|
"tn": 0,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"forget_correction": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 0,
|
|
"unknown_episodes": 2000,
|
|
"top1": null,
|
|
"recall1": null,
|
|
"recall3": null,
|
|
"recall5": null,
|
|
"recall8": null,
|
|
"evidence_count_in_top3": null,
|
|
"hit3": null,
|
|
"mrr": null,
|
|
"ndcg3": null,
|
|
"mean_first_positive_rank": null,
|
|
"all_evidence_in_top3": null,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": null,
|
|
"hop_under_prediction": null,
|
|
"hop_over_prediction": null,
|
|
"mean_score_margin": null,
|
|
"median_score_margin": null,
|
|
"negative_margin_rate": null,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 2000
|
|
},
|
|
"known_mean_need_prob": null,
|
|
"unknown_mean_need_prob": 1.275862421441728e-10,
|
|
"thr0.50": {
|
|
"need_precision": null,
|
|
"need_recall": null,
|
|
"need_f1": null,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": null,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 0,
|
|
"tn": 2000,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"long_context": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.9795,
|
|
"recall1": 0.9795,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9896666666666666,
|
|
"ndcg3": 0.9923685950714292,
|
|
"mean_first_positive_rank": 1.021,
|
|
"all_evidence_in_top3": 1.0,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 60.150181606799364,
|
|
"median_score_margin": 68.01980018615723,
|
|
"negative_margin_rate": 0.0205,
|
|
"hop_confusion": {
|
|
"true_1_pred_1": 2000
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": null,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": null,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": null,
|
|
"counts": {
|
|
"tp": 2000,
|
|
"tn": 0,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"multi_hop": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.9885,
|
|
"recall1": 0.49425,
|
|
"recall3": 0.9605,
|
|
"recall5": 0.99825,
|
|
"recall8": 0.999,
|
|
"evidence_count_in_top3": 1.921,
|
|
"hit3": 1.0,
|
|
"mrr": 0.99425,
|
|
"ndcg3": 0.9452009954703022,
|
|
"mean_first_positive_rank": 1.0115,
|
|
"all_evidence_in_top3": 0.921,
|
|
"multi_positive_episodes": 2000,
|
|
"all_evidence_in_top3_multi": 0.921,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 55.6439196386747,
|
|
"median_score_margin": 62.50904202461243,
|
|
"negative_margin_rate": 0.0115,
|
|
"hop_confusion": {
|
|
"true_2_pred_2": 2000
|
|
},
|
|
"known_mean_need_prob": 0.9999999997615814,
|
|
"unknown_mean_need_prob": null,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": null,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": null,
|
|
"counts": {
|
|
"tp": 2000,
|
|
"tn": 0,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"paraphrase": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.995,
|
|
"recall1": 0.995,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9975,
|
|
"ndcg3": 0.9981546487678573,
|
|
"mean_first_positive_rank": 1.005,
|
|
"all_evidence_in_top3": 1.0,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 48.13904330630228,
|
|
"median_score_margin": 49.25835180282593,
|
|
"negative_margin_rate": 0.005,
|
|
"hop_confusion": {
|
|
"true_1_pred_1": 2000
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": null,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": null,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": null,
|
|
"counts": {
|
|
"tp": 2000,
|
|
"tn": 0,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"random_position": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.9805,
|
|
"recall1": 0.9805,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9901666666666666,
|
|
"ndcg3": 0.9927376653178577,
|
|
"mean_first_positive_rank": 1.02,
|
|
"all_evidence_in_top3": 1.0,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 50.85701001610607,
|
|
"median_score_margin": 54.9089093208313,
|
|
"negative_margin_rate": 0.0195,
|
|
"hop_confusion": {
|
|
"true_1_pred_1": 2000
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": null,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": null,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": null,
|
|
"counts": {
|
|
"tp": 2000,
|
|
"tn": 0,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"single_fact": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.9745,
|
|
"recall1": 0.9745,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9871666666666666,
|
|
"ndcg3": 0.9905232438392865,
|
|
"mean_first_positive_rank": 1.026,
|
|
"all_evidence_in_top3": 1.0,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 57.84846639241278,
|
|
"median_score_margin": 63.093711376190186,
|
|
"negative_margin_rate": 0.0255,
|
|
"hop_confusion": {
|
|
"true_1_pred_1": 2000
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": null,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": null,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": null,
|
|
"counts": {
|
|
"tp": 2000,
|
|
"tn": 0,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"unknown": {
|
|
"episodes": 1920,
|
|
"answerable_episodes": 1814,
|
|
"unknown_episodes": 106,
|
|
"top1": 0.5898566703417861,
|
|
"recall1": 0.5898566703417861,
|
|
"recall3": 0.6896361631753032,
|
|
"recall5": 0.7464167585446527,
|
|
"recall8": 0.8676957001102535,
|
|
"evidence_count_in_top3": 0.6896361631753032,
|
|
"hit3": 0.6896361631753032,
|
|
"mrr": 0.6759606339151697,
|
|
"ndcg3": 0.6479024598421029,
|
|
"mean_first_positive_rank": 3.383131201764057,
|
|
"all_evidence_in_top3": 0.6896361631753032,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 0.20369631854576997,
|
|
"median_score_margin": 6.827788352966309,
|
|
"negative_margin_rate": 0.4101433296582139,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 106,
|
|
"true_0_pred_1": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 1814
|
|
},
|
|
"known_mean_need_prob": 0.9999999848195447,
|
|
"unknown_mean_need_prob": 0.00018733273071117103,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 1814,
|
|
"tn": 106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"unknown_abstention": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 0,
|
|
"unknown_episodes": 2000,
|
|
"top1": null,
|
|
"recall1": null,
|
|
"recall3": null,
|
|
"recall5": null,
|
|
"recall8": null,
|
|
"evidence_count_in_top3": null,
|
|
"hit3": null,
|
|
"mrr": null,
|
|
"ndcg3": null,
|
|
"mean_first_positive_rank": null,
|
|
"all_evidence_in_top3": null,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": null,
|
|
"hop_under_prediction": null,
|
|
"hop_over_prediction": null,
|
|
"mean_score_margin": null,
|
|
"median_score_margin": null,
|
|
"negative_margin_rate": null,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 2000
|
|
},
|
|
"known_mean_need_prob": null,
|
|
"unknown_mean_need_prob": 3.3919047850488803e-15,
|
|
"thr0.50": {
|
|
"need_precision": null,
|
|
"need_recall": null,
|
|
"need_f1": null,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": null,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 0,
|
|
"tn": 2000,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
}
|
|
}
|
|
},
|
|
"REPLAY-128 final": {
|
|
"checkpoint": "checkpoints\\router_replay_v7_v2_128\\memory_router_v2.pt",
|
|
"router": {
|
|
"kind": "router_v2",
|
|
"step": null,
|
|
"arch_config": {
|
|
"arch": "router_v2",
|
|
"hidden_size": 2560,
|
|
"router_dim": 128,
|
|
"num_heads": 8,
|
|
"max_hops": 3
|
|
},
|
|
"parameters": 2037774,
|
|
"checkpoint_bytes": 8156921
|
|
},
|
|
"storage": {
|
|
"address_bytes_per_record": 512,
|
|
"address_mb_per_1m_records": 512,
|
|
"checkpoint_bytes": 8156921
|
|
},
|
|
"latency": {
|
|
"cuda": {
|
|
"single_query_latency_ms_mean": 0.9342146668738375,
|
|
"single_query_latency_ms_p50": 0.8827999990899116,
|
|
"single_query_latency_ms_p95": 1.4861000017845072,
|
|
"queries_per_second": 1070.417790962649,
|
|
"samples": 150,
|
|
"batched_qps_64": 70303.40306999629,
|
|
"batched_latency_ms_64": 0.014224062510947988,
|
|
"batched_qps_256": 244326.0991071313,
|
|
"batched_latency_ms_256": 0.004092890623041967
|
|
},
|
|
"cpu": {
|
|
"single_query_latency_ms_mean": 0.3855180000149024,
|
|
"single_query_latency_ms_p50": 0.36720000207424164,
|
|
"single_query_latency_ms_p95": 0.5253999988781288,
|
|
"queries_per_second": 2593.912605796213,
|
|
"samples": 150,
|
|
"batched_qps_64": 12253.306479886014,
|
|
"batched_latency_ms_64": 0.08161062498857063,
|
|
"batched_qps_256": 12008.169307919601,
|
|
"batched_latency_ms_256": 0.08327664062335316
|
|
}
|
|
},
|
|
"metrics": {
|
|
"episodes": 21920,
|
|
"answerable_episodes": 17814,
|
|
"unknown_episodes": 4106,
|
|
"top1": 0.9413944088918828,
|
|
"recall1": 0.8857920736499383,
|
|
"recall3": 0.9647748961490962,
|
|
"recall5": 0.9714550353654429,
|
|
"recall8": 0.9804086673402942,
|
|
"evidence_count_in_top3": 1.0744919726058157,
|
|
"hit3": 0.9673290670259347,
|
|
"mrr": 0.957689277532011,
|
|
"ndcg3": 0.9536787031890833,
|
|
"mean_first_positive_rank": 1.34001347254968,
|
|
"all_evidence_in_top3": 0.9622207252722578,
|
|
"multi_positive_episodes": 2000,
|
|
"all_evidence_in_top3_multi": 0.9545,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 40.14557690149532,
|
|
"median_score_margin": 45.34479379653931,
|
|
"negative_margin_rate": 0.05860559110811721,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 4106,
|
|
"true_0_pred_1": 0,
|
|
"true_0_pred_2": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 15814,
|
|
"true_1_pred_2": 0,
|
|
"true_2_pred_0": 0,
|
|
"true_2_pred_1": 0,
|
|
"true_2_pred_2": 2000
|
|
},
|
|
"known_mean_need_prob": 0.9999999984274064,
|
|
"unknown_mean_need_prob": 4.836221294362452e-06,
|
|
"thr0.30": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.40": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.60": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.70": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
},
|
|
"thr0.80": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 17814,
|
|
"tn": 4106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"by_family": {
|
|
"benchmark_qa": {
|
|
"episodes": 64,
|
|
"answerable_episodes": 0,
|
|
"unknown_episodes": 64,
|
|
"top1": null,
|
|
"recall1": null,
|
|
"recall3": null,
|
|
"recall5": null,
|
|
"recall8": null,
|
|
"evidence_count_in_top3": null,
|
|
"hit3": null,
|
|
"mrr": null,
|
|
"ndcg3": null,
|
|
"mean_first_positive_rank": null,
|
|
"all_evidence_in_top3": null,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": null,
|
|
"hop_under_prediction": null,
|
|
"hop_over_prediction": null,
|
|
"mean_score_margin": null,
|
|
"median_score_margin": null,
|
|
"negative_margin_rate": null,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 64
|
|
},
|
|
"known_mean_need_prob": null,
|
|
"unknown_mean_need_prob": 4.8406160974828055e-06,
|
|
"thr0.50": {
|
|
"need_precision": null,
|
|
"need_recall": null,
|
|
"need_f1": null,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": null,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 0,
|
|
"tn": 64,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"mega_validation": {
|
|
"episodes": 20000,
|
|
"answerable_episodes": 16000,
|
|
"unknown_episodes": 4000,
|
|
"top1": 0.98225,
|
|
"recall1": 0.92034375,
|
|
"recall3": 0.99715625,
|
|
"recall5": 0.99959375,
|
|
"recall8": 0.9996875,
|
|
"evidence_count_in_top3": 1.1193125,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9911041666666667,
|
|
"ndcg3": 0.9893498205729669,
|
|
"mean_first_positive_rank": 1.017875,
|
|
"all_evidence_in_top3": 0.9943125,
|
|
"multi_positive_episodes": 2000,
|
|
"all_evidence_in_top3_multi": 0.9545,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 44.73340179644042,
|
|
"median_score_margin": 47.92497134208679,
|
|
"negative_margin_rate": 0.01775,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 4000,
|
|
"true_0_pred_1": 0,
|
|
"true_0_pred_2": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 14000,
|
|
"true_1_pred_2": 0,
|
|
"true_2_pred_0": 0,
|
|
"true_2_pred_1": 0,
|
|
"true_2_pred_2": 2000
|
|
},
|
|
"known_mean_need_prob": 0.9999999999701977,
|
|
"unknown_mean_need_prob": 6.379481702447893e-11,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 16000,
|
|
"tn": 4000,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"memory_policy": {
|
|
"episodes": 1728,
|
|
"answerable_episodes": 1707,
|
|
"unknown_episodes": 21,
|
|
"top1": 0.5875805506736965,
|
|
"recall1": 0.5875805506736965,
|
|
"recall3": 0.6731107205623902,
|
|
"recall5": 0.7106033977738723,
|
|
"recall8": 0.7996485061511424,
|
|
"evidence_count_in_top3": 0.6731107205623902,
|
|
"hit3": 0.6731107205623902,
|
|
"mrr": 0.6633131807141506,
|
|
"ndcg3": 0.6380925044585338,
|
|
"mean_first_positive_rank": 4.295254833040421,
|
|
"all_evidence_in_top3": 0.6731107205623902,
|
|
"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": -0.2721888081208264,
|
|
"median_score_margin": 8.391286849975586,
|
|
"negative_margin_rate": 0.4124194493263035,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 21,
|
|
"true_0_pred_1": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 1707
|
|
},
|
|
"known_mean_need_prob": 0.9999999838679872,
|
|
"unknown_mean_need_prob": 0.0004654159529796483,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 1707,
|
|
"tn": 21,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"native_memory": {
|
|
"episodes": 128,
|
|
"answerable_episodes": 107,
|
|
"unknown_episodes": 21,
|
|
"top1": 0.4766355140186916,
|
|
"recall1": 0.4766355140186916,
|
|
"recall3": 0.7757009345794392,
|
|
"recall5": 0.9252336448598131,
|
|
"recall8": 0.9813084112149533,
|
|
"evidence_count_in_top3": 0.7757009345794392,
|
|
"hit3": 0.7757009345794392,
|
|
"mrr": 0.6573319982198488,
|
|
"ndcg3": 0.6543120031041451,
|
|
"mean_first_positive_rank": 2.364485981308411,
|
|
"all_evidence_in_top3": 0.7757009345794392,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": -1.0887432181946586,
|
|
"median_score_margin": -0.6999244689941406,
|
|
"negative_margin_rate": 0.5233644859813084,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 21,
|
|
"true_0_pred_1": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 107
|
|
},
|
|
"known_mean_need_prob": 1.0,
|
|
"unknown_mean_need_prob": 0.0004654159529796483,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 107,
|
|
"tn": 21,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
}
|
|
},
|
|
"by_category": {
|
|
"conflict_update": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.991,
|
|
"recall1": 0.991,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9955,
|
|
"ndcg3": 0.9966783677821431,
|
|
"mean_first_positive_rank": 1.009,
|
|
"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": 42.24227243242413,
|
|
"median_score_margin": 42.77137851715088,
|
|
"negative_margin_rate": 0.009,
|
|
"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": 0.973,
|
|
"recall1": 0.973,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9865,
|
|
"ndcg3": 0.9900351033464293,
|
|
"mean_first_positive_rank": 1.027,
|
|
"all_evidence_in_top3": 1.0,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 50.82861385926604,
|
|
"median_score_margin": 54.98523235321045,
|
|
"negative_margin_rate": 0.027,
|
|
"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": 0.983,
|
|
"recall1": 0.983,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9915,
|
|
"ndcg3": 0.9937258058107148,
|
|
"mean_first_positive_rank": 1.017,
|
|
"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": 51.79758185644448,
|
|
"median_score_margin": 55.65491724014282,
|
|
"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
|
|
}
|
|
}
|
|
},
|
|
"forget_correction": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 0,
|
|
"unknown_episodes": 2000,
|
|
"top1": null,
|
|
"recall1": null,
|
|
"recall3": null,
|
|
"recall5": null,
|
|
"recall8": null,
|
|
"evidence_count_in_top3": null,
|
|
"hit3": null,
|
|
"mrr": null,
|
|
"ndcg3": null,
|
|
"mean_first_positive_rank": null,
|
|
"all_evidence_in_top3": null,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": null,
|
|
"hop_under_prediction": null,
|
|
"hop_over_prediction": null,
|
|
"mean_score_margin": null,
|
|
"median_score_margin": null,
|
|
"negative_margin_rate": null,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 2000
|
|
},
|
|
"known_mean_need_prob": null,
|
|
"unknown_mean_need_prob": 1.275862421441728e-10,
|
|
"thr0.50": {
|
|
"need_precision": null,
|
|
"need_recall": null,
|
|
"need_f1": null,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": null,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 0,
|
|
"tn": 2000,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"long_context": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 2000,
|
|
"unknown_episodes": 0,
|
|
"top1": 0.978,
|
|
"recall1": 0.978,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9889166666666667,
|
|
"ndcg3": 0.9918149897017863,
|
|
"mean_first_positive_rank": 1.0225,
|
|
"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": 47.034401270009575,
|
|
"median_score_margin": 50.28937745094299,
|
|
"negative_margin_rate": 0.022,
|
|
"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": 0.9905,
|
|
"recall1": 0.49525,
|
|
"recall3": 0.97725,
|
|
"recall5": 0.99675,
|
|
"recall8": 0.9975,
|
|
"evidence_count_in_top3": 1.9545,
|
|
"hit3": 1.0,
|
|
"mrr": 0.99525,
|
|
"ndcg3": 0.963831301989089,
|
|
"mean_first_positive_rank": 1.0095,
|
|
"all_evidence_in_top3": 0.9545,
|
|
"multi_positive_episodes": 2000,
|
|
"all_evidence_in_top3_multi": 0.9545,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": 1.0,
|
|
"hop_under_prediction": 0.0,
|
|
"hop_over_prediction": 0.0,
|
|
"mean_score_margin": 46.103771095309405,
|
|
"median_score_margin": 52.25391101837158,
|
|
"negative_margin_rate": 0.0095,
|
|
"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": 0.994,
|
|
"recall1": 0.994,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.997,
|
|
"ndcg3": 0.9977855785214288,
|
|
"mean_first_positive_rank": 1.006,
|
|
"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": 36.59333083114028,
|
|
"median_score_margin": 35.444119691848755,
|
|
"negative_margin_rate": 0.006,
|
|
"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": 0.978,
|
|
"recall1": 0.978,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.9889166666666667,
|
|
"ndcg3": 0.9918149897017863,
|
|
"mean_first_positive_rank": 1.0225,
|
|
"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": 35.780669214252384,
|
|
"median_score_margin": 35.064393758773804,
|
|
"negative_margin_rate": 0.022,
|
|
"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": 0.9705,
|
|
"recall1": 0.9705,
|
|
"recall3": 1.0,
|
|
"recall5": 1.0,
|
|
"recall8": 1.0,
|
|
"evidence_count_in_top3": 1.0,
|
|
"hit3": 1.0,
|
|
"mrr": 0.98525,
|
|
"ndcg3": 0.989112427730358,
|
|
"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": 47.48657381267706,
|
|
"median_score_margin": 51.48151856660843,
|
|
"negative_margin_rate": 0.0295,
|
|
"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": 0.5810363836824697,
|
|
"recall1": 0.5810363836824697,
|
|
"recall3": 0.679162072767365,
|
|
"recall5": 0.7232635060639471,
|
|
"recall8": 0.8103638368246968,
|
|
"evidence_count_in_top3": 0.679162072767365,
|
|
"hit3": 0.679162072767365,
|
|
"mrr": 0.662960376675071,
|
|
"ndcg3": 0.6390492224051051,
|
|
"mean_first_positive_rank": 4.181367144432194,
|
|
"all_evidence_in_top3": 0.679162072767365,
|
|
"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": -0.32035381466873164,
|
|
"median_score_margin": 7.441610336303711,
|
|
"negative_margin_rate": 0.4189636163175303,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 106,
|
|
"true_0_pred_1": 0,
|
|
"true_1_pred_0": 0,
|
|
"true_1_pred_1": 1814
|
|
},
|
|
"known_mean_need_prob": 0.9999999848195447,
|
|
"unknown_mean_need_prob": 0.00018733273071117103,
|
|
"thr0.50": {
|
|
"need_precision": 1.0,
|
|
"need_recall": 1.0,
|
|
"need_f1": 1.0,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": 0.0,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 1814,
|
|
"tn": 106,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
},
|
|
"unknown_abstention": {
|
|
"episodes": 2000,
|
|
"answerable_episodes": 0,
|
|
"unknown_episodes": 2000,
|
|
"top1": null,
|
|
"recall1": null,
|
|
"recall3": null,
|
|
"recall5": null,
|
|
"recall8": null,
|
|
"evidence_count_in_top3": null,
|
|
"hit3": null,
|
|
"mrr": null,
|
|
"ndcg3": null,
|
|
"mean_first_positive_rank": null,
|
|
"all_evidence_in_top3": null,
|
|
"multi_positive_episodes": 0,
|
|
"all_evidence_in_top3_multi": null,
|
|
"hop_accuracy": 1.0,
|
|
"hop_accuracy_answerable": null,
|
|
"hop_under_prediction": null,
|
|
"hop_over_prediction": null,
|
|
"mean_score_margin": null,
|
|
"median_score_margin": null,
|
|
"negative_margin_rate": null,
|
|
"hop_confusion": {
|
|
"true_0_pred_0": 2000
|
|
},
|
|
"known_mean_need_prob": null,
|
|
"unknown_mean_need_prob": 3.3919047850488803e-15,
|
|
"thr0.50": {
|
|
"need_precision": null,
|
|
"need_recall": null,
|
|
"need_f1": null,
|
|
"specificity_unknown_refusal": 1.0,
|
|
"abstention_accuracy": 1.0,
|
|
"known_question_refusal_rate": null,
|
|
"unknown_question_read_rate": 0.0,
|
|
"counts": {
|
|
"tp": 0,
|
|
"tn": 2000,
|
|
"fp": 0,
|
|
"fn": 0
|
|
}
|
|
}
|
|
}
|
|
}
|
|
}
|
|
} |