andykrik's picture
Upload metrics.json with huggingface_hub
c7952b5 verified
[
{
"epoch": 1,
"train_loss": 5.157754315270318,
"train_tool_retrieval_accuracy": 0.8961538672447205,
"train_parent_retrieval_accuracy": 0.7846153974533081,
"val_tool_retrieval_accuracy": 0.5192307829856873,
"val_parent_retrieval_accuracy": 0.692307710647583,
"val_tool_classification_accuracy": 0.0,
"val_parent_classification_accuracy": 0.3076923076923077,
"learning_rate": 8.181818181818183e-06,
"epoch_duration_seconds": 7.761668143000861,
"best_val_tool_retrieval_accuracy": 0.5192307829856873,
"best_epoch": 1,
"train_parent_cross_entropy": 2.3311243984434338,
"train_child_cross_entropy": 2.625902043448554,
"train_parent_alignment": 0.9960288206736246,
"train_child_alignment": 1.0112511979209051,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.5192307692307693,
"validation/tool/silhouette_score": -0.09840114414691925,
"validation/tool/global_near_overlap_rate": 0.7692307692307693,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.6923076923076923,
"validation/parent/silhouette_score": -0.07535062730312347,
"validation/parent/global_near_overlap_rate": 0.6153846153846154,
"validation/parent/top_pair/label_a": "File System",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.5,
"validation/parent/top_pair/centroid_cosine": 0.896780788898468
},
{
"epoch": 2,
"train_loss": 4.649685700734456,
"train_tool_retrieval_accuracy": 0.8730769157409668,
"train_parent_retrieval_accuracy": 0.7769230604171753,
"val_tool_retrieval_accuracy": 0.5192307829856873,
"val_parent_retrieval_accuracy": 0.7307692170143127,
"val_tool_classification_accuracy": 0.038461538461538464,
"val_parent_classification_accuracy": 0.40384615384615385,
"learning_rate": 1.6363636363636366e-05,
"epoch_duration_seconds": 1.7147771270028898,
"best_val_tool_retrieval_accuracy": 0.5192307829856873,
"best_epoch": 1,
"train_parent_cross_entropy": 1.930469234784444,
"train_child_cross_entropy": 2.5211688147650824,
"train_parent_alignment": 0.9734141098128425,
"train_child_alignment": 1.0070623556772869,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.5192307692307693,
"validation/tool/silhouette_score": -0.10537859797477722,
"validation/tool/global_near_overlap_rate": 0.7884615384615384,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.7307692307692307,
"validation/parent/silhouette_score": -0.07874973863363266,
"validation/parent/global_near_overlap_rate": 0.6538461538461539,
"validation/parent/top_pair/label_a": "File System",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.44999998807907104,
"validation/parent/top_pair/centroid_cosine": 0.9004660844802856
},
{
"epoch": 3,
"train_loss": 3.500072797139486,
"train_tool_retrieval_accuracy": 0.8769230842590332,
"train_parent_retrieval_accuracy": 0.7846153974533081,
"val_tool_retrieval_accuracy": 0.5,
"val_parent_retrieval_accuracy": 0.75,
"val_tool_classification_accuracy": 0.07692307692307693,
"val_parent_classification_accuracy": 0.5384615384615384,
"learning_rate": 1.950738916256158e-05,
"epoch_duration_seconds": 1.303288903000066,
"best_val_tool_retrieval_accuracy": 0.5192307829856873,
"best_epoch": 1,
"train_parent_cross_entropy": 1.39888596534729,
"train_child_cross_entropy": 1.9089170164532132,
"train_parent_alignment": 0.9373632669448853,
"train_child_alignment": 0.9853348599539863,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.5,
"validation/tool/silhouette_score": -0.10424783825874329,
"validation/tool/global_near_overlap_rate": 0.8076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.75,
"validation/parent/silhouette_score": -0.08332549035549164,
"validation/parent/global_near_overlap_rate": 0.6346153846153846,
"validation/parent/top_pair/label_a": "File System",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.45000001788139343,
"validation/parent/top_pair/centroid_cosine": 0.8934528231620789
},
{
"epoch": 4,
"train_loss": 2.730307208167182,
"train_tool_retrieval_accuracy": 0.8961538672447205,
"train_parent_retrieval_accuracy": 0.807692289352417,
"val_tool_retrieval_accuracy": 0.5384615659713745,
"val_parent_retrieval_accuracy": 0.7692307829856873,
"val_tool_classification_accuracy": 0.17307692307692307,
"val_parent_classification_accuracy": 0.7115384615384616,
"learning_rate": 1.8620689655172415e-05,
"epoch_duration_seconds": 1.3154947980001452,
"best_val_tool_retrieval_accuracy": 0.5384615659713745,
"best_epoch": 4,
"train_parent_cross_entropy": 1.0005325476328533,
"train_child_cross_entropy": 1.5424939393997192,
"train_parent_alignment": 0.9058522052235074,
"train_child_alignment": 0.9669552577866448,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.5384615384615384,
"validation/tool/silhouette_score": -0.09466496855020523,
"validation/tool/global_near_overlap_rate": 0.7884615384615384,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.7692307692307693,
"validation/parent/silhouette_score": -0.0762428417801857,
"validation/parent/global_near_overlap_rate": 0.5769230769230769,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Information Retrieval",
"validation/parent/top_pair/mutual_overlap": 0.44999998807907104,
"validation/parent/top_pair/centroid_cosine": 0.8568636178970337
},
{
"epoch": 5,
"train_loss": 2.2074906561109753,
"train_tool_retrieval_accuracy": 0.9076923131942749,
"train_parent_retrieval_accuracy": 0.8461538553237915,
"val_tool_retrieval_accuracy": 0.5961538553237915,
"val_parent_retrieval_accuracy": 0.7692307829856873,
"val_tool_classification_accuracy": 0.3076923076923077,
"val_parent_classification_accuracy": 0.75,
"learning_rate": 1.7733990147783254e-05,
"epoch_duration_seconds": 1.333961725998961,
"best_val_tool_retrieval_accuracy": 0.5961538553237915,
"best_epoch": 5,
"train_parent_cross_entropy": 0.6690340472592248,
"train_child_cross_entropy": 1.3546243641111586,
"train_parent_alignment": 0.8832898471090529,
"train_child_alignment": 0.9550319446457757,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.5961538461538461,
"validation/tool/silhouette_score": -0.0741507038474083,
"validation/tool/global_near_overlap_rate": 0.7307692307692307,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.7692307692307693,
"validation/parent/silhouette_score": -0.06206924095749855,
"validation/parent/global_near_overlap_rate": 0.38461538461538464,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Information Retrieval",
"validation/parent/top_pair/mutual_overlap": 0.3499999940395355,
"validation/parent/top_pair/centroid_cosine": 0.8201030492782593
},
{
"epoch": 6,
"train_loss": 1.5056148568789165,
"train_tool_retrieval_accuracy": 0.9346153736114502,
"train_parent_retrieval_accuracy": 0.8807692527770996,
"val_tool_retrieval_accuracy": 0.6538461446762085,
"val_parent_retrieval_accuracy": 0.807692289352417,
"val_tool_classification_accuracy": 0.4423076923076923,
"val_parent_classification_accuracy": 0.7884615384615384,
"learning_rate": 1.684729064039409e-05,
"epoch_duration_seconds": 2.3185505299989018,
"best_val_tool_retrieval_accuracy": 0.6538461446762085,
"best_epoch": 6,
"train_parent_cross_entropy": 0.46070420576466453,
"train_child_cross_entropy": 0.8664655850993263,
"train_parent_alignment": 0.8618110550774468,
"train_child_alignment": 0.9226394295692444,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.6538461538461539,
"validation/tool/silhouette_score": -0.03729083389043808,
"validation/tool/global_near_overlap_rate": 0.6730769230769231,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.8076923076923077,
"validation/parent/silhouette_score": -0.03389709070324898,
"validation/parent/global_near_overlap_rate": 0.3269230769230769,
"validation/parent/top_pair/label_a": "Information Retrieval",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.21666666865348816,
"validation/parent/top_pair/centroid_cosine": 0.8001468181610107
},
{
"epoch": 7,
"train_loss": 1.2175099054972331,
"train_tool_retrieval_accuracy": 0.9461538195610046,
"train_parent_retrieval_accuracy": 0.9115384817123413,
"val_tool_retrieval_accuracy": 0.7115384340286255,
"val_parent_retrieval_accuracy": 0.8269230723381042,
"val_tool_classification_accuracy": 0.5,
"val_parent_classification_accuracy": 0.8076923076923077,
"learning_rate": 1.5960591133004928e-05,
"epoch_duration_seconds": 1.3010024789982708,
"best_val_tool_retrieval_accuracy": 0.7115384340286255,
"best_epoch": 7,
"train_parent_cross_entropy": 0.34940624402629006,
"train_child_cross_entropy": 0.6940820945633782,
"train_parent_alignment": 0.8370489080746969,
"train_child_alignment": 0.9031666649712456,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7115384615384616,
"validation/tool/silhouette_score": -0.012219934724271297,
"validation/tool/global_near_overlap_rate": 0.6346153846153846,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.8269230769230769,
"validation/parent/silhouette_score": -0.012261098250746727,
"validation/parent/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.20000000298023224,
"validation/parent/top_pair/centroid_cosine": 0.7784289121627808
},
{
"epoch": 8,
"train_loss": 0.92488608095381,
"train_tool_retrieval_accuracy": 0.9576923251152039,
"train_parent_retrieval_accuracy": 0.9230769276618958,
"val_tool_retrieval_accuracy": 0.6730769276618958,
"val_parent_retrieval_accuracy": 0.8269230723381042,
"val_tool_classification_accuracy": 0.5961538461538461,
"val_parent_classification_accuracy": 0.8461538461538461,
"learning_rate": 1.5073891625615764e-05,
"epoch_duration_seconds": 1.3682630530020106,
"best_val_tool_retrieval_accuracy": 0.7115384340286255,
"best_epoch": 7,
"train_parent_cross_entropy": 0.23755988809797499,
"train_child_cross_entropy": 0.5170698364575704,
"train_parent_alignment": 0.8208589951197306,
"train_child_alignment": 0.8817045291264852,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.6730769230769231,
"validation/tool/silhouette_score": 0.004278222564607859,
"validation/tool/global_near_overlap_rate": 0.5769230769230769,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.8269230769230769,
"validation/parent/silhouette_score": 0.0034682746045291424,
"validation/parent/global_near_overlap_rate": 0.2692307692307692,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.20000000298023224,
"validation/parent/top_pair/centroid_cosine": 0.7186568975448608
},
{
"epoch": 9,
"train_loss": 0.7474626302719116,
"train_tool_retrieval_accuracy": 0.9653846025466919,
"train_parent_retrieval_accuracy": 0.9384615421295166,
"val_tool_retrieval_accuracy": 0.6730769276618958,
"val_parent_retrieval_accuracy": 0.8461538553237915,
"val_tool_classification_accuracy": 0.6346153846153846,
"val_parent_classification_accuracy": 0.8461538461538461,
"learning_rate": 1.4187192118226602e-05,
"epoch_duration_seconds": 2.020530112000415,
"best_val_tool_retrieval_accuracy": 0.7115384340286255,
"best_epoch": 7,
"train_parent_cross_entropy": 0.19186763796541426,
"train_child_cross_entropy": 0.38965080512894523,
"train_parent_alignment": 0.800107995669047,
"train_child_alignment": 0.8593339323997498,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.6730769230769231,
"validation/tool/silhouette_score": 0.02403559535741806,
"validation/tool/global_near_overlap_rate": 0.5384615384615384,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.8461538461538461,
"validation/parent/silhouette_score": 0.017015840858221054,
"validation/parent/global_near_overlap_rate": 0.25,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.15000000596046448,
"validation/parent/top_pair/centroid_cosine": 0.6582691669464111
},
{
"epoch": 10,
"train_loss": 0.560799671543969,
"train_tool_retrieval_accuracy": 0.9730769395828247,
"train_parent_retrieval_accuracy": 0.949999988079071,
"val_tool_retrieval_accuracy": 0.692307710647583,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.6538461538461539,
"val_parent_classification_accuracy": 0.8461538461538461,
"learning_rate": 1.330049261083744e-05,
"epoch_duration_seconds": 1.3082187740001245,
"best_val_tool_retrieval_accuracy": 0.7115384340286255,
"best_epoch": 7,
"train_parent_cross_entropy": 0.11244031579958068,
"train_child_cross_entropy": 0.28610827691025204,
"train_parent_alignment": 0.7726466457049052,
"train_child_alignment": 0.8498640192879571,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.6923076923076923,
"validation/tool/silhouette_score": 0.03721160441637039,
"validation/tool/global_near_overlap_rate": 0.4423076923076923,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.03269634023308754,
"validation/parent/global_near_overlap_rate": 0.23076923076923078,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.15000000596046448,
"validation/parent/top_pair/centroid_cosine": 0.6069332361221313
},
{
"epoch": 11,
"train_loss": 0.4488269024425083,
"train_tool_retrieval_accuracy": 0.9769230484962463,
"train_parent_retrieval_accuracy": 0.9653846025466919,
"val_tool_retrieval_accuracy": 0.7307692170143127,
"val_parent_retrieval_accuracy": 0.9230769276618958,
"val_tool_classification_accuracy": 0.6923076923076923,
"val_parent_classification_accuracy": 0.8653846153846154,
"learning_rate": 1.2413793103448277e-05,
"epoch_duration_seconds": 1.3191320069963695,
"best_val_tool_retrieval_accuracy": 0.7307692170143127,
"best_epoch": 11,
"train_parent_cross_entropy": 0.09428231045603752,
"train_child_cross_entropy": 0.19627481864558327,
"train_parent_alignment": 0.7582567930221558,
"train_child_alignment": 0.8244408766428629,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7307692307692307,
"validation/tool/silhouette_score": 0.04294625297188759,
"validation/tool/global_near_overlap_rate": 0.40384615384615385,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9230769230769231,
"validation/parent/silhouette_score": 0.04077092185616493,
"validation/parent/global_near_overlap_rate": 0.25,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.15000000596046448,
"validation/parent/top_pair/centroid_cosine": 0.5526243448257446
},
{
"epoch": 12,
"train_loss": 0.38162952330377364,
"train_tool_retrieval_accuracy": 0.9846153855323792,
"train_parent_retrieval_accuracy": 0.9730769395828247,
"val_tool_retrieval_accuracy": 0.7692307829856873,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.6923076923076923,
"val_parent_classification_accuracy": 0.8653846153846154,
"learning_rate": 1.1527093596059114e-05,
"epoch_duration_seconds": 1.3753753389974008,
"best_val_tool_retrieval_accuracy": 0.7692307829856873,
"best_epoch": 12,
"train_parent_cross_entropy": 0.07771683939629132,
"train_child_cross_entropy": 0.1482034358713362,
"train_parent_alignment": 0.7491327060593499,
"train_child_alignment": 0.807959775129954,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7692307692307693,
"validation/tool/silhouette_score": 0.05038919299840927,
"validation/tool/global_near_overlap_rate": 0.38461538461538464,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.04405605047941208,
"validation/parent/global_near_overlap_rate": 0.23076923076923078,
"validation/parent/top_pair/label_a": "Information Retrieval",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.13333334028720856,
"validation/parent/top_pair/centroid_cosine": 0.3341650366783142
},
{
"epoch": 13,
"train_loss": 0.34940913650724625,
"train_tool_retrieval_accuracy": 0.9846153855323792,
"train_parent_retrieval_accuracy": 0.9807692170143127,
"val_tool_retrieval_accuracy": 0.7692307829856873,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.7115384615384616,
"val_parent_classification_accuracy": 0.8653846153846154,
"learning_rate": 1.0640394088669953e-05,
"epoch_duration_seconds": 1.9898250920014107,
"best_val_tool_retrieval_accuracy": 0.7692307829856873,
"best_epoch": 12,
"train_parent_cross_entropy": 0.05860658983389536,
"train_child_cross_entropy": 0.1373151573869917,
"train_parent_alignment": 0.7339069644610087,
"train_child_alignment": 0.8009669250912137,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7692307692307693,
"validation/tool/silhouette_score": 0.06007323041558266,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.04259569197893143,
"validation/parent/global_near_overlap_rate": 0.25,
"validation/parent/top_pair/label_a": "Information Retrieval",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.13333334028720856,
"validation/parent/top_pair/centroid_cosine": 0.2915511727333069
},
{
"epoch": 14,
"train_loss": 0.3008048352268007,
"train_tool_retrieval_accuracy": 0.9884615540504456,
"train_parent_retrieval_accuracy": 0.9923076629638672,
"val_tool_retrieval_accuracy": 0.75,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.7115384615384616,
"val_parent_classification_accuracy": 0.8461538461538461,
"learning_rate": 9.75369458128079e-06,
"epoch_duration_seconds": 1.3389712629941641,
"best_val_tool_retrieval_accuracy": 0.7692307829856873,
"best_epoch": 12,
"train_parent_cross_entropy": 0.04300534518228637,
"train_child_cross_entropy": 0.10681118857529429,
"train_parent_alignment": 0.7165413432651095,
"train_child_alignment": 0.7933415704303317,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.75,
"validation/tool/silhouette_score": 0.063718281686306,
"validation/tool/global_near_overlap_rate": 0.3269230769230769,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.05269314721226692,
"validation/parent/global_near_overlap_rate": 0.21153846153846154,
"validation/parent/top_pair/label_a": "Information Retrieval",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.13333334028720856,
"validation/parent/top_pair/centroid_cosine": 0.25321274995803833
},
{
"epoch": 15,
"train_loss": 0.2826983431975047,
"train_tool_retrieval_accuracy": 0.9923076629638672,
"train_parent_retrieval_accuracy": 0.9961538314819336,
"val_tool_retrieval_accuracy": 0.7692307829856873,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.7115384615384616,
"val_parent_classification_accuracy": 0.8653846153846154,
"learning_rate": 8.866995073891627e-06,
"epoch_duration_seconds": 1.3003545039973687,
"best_val_tool_retrieval_accuracy": 0.7692307829856873,
"best_epoch": 12,
"train_parent_cross_entropy": 0.04354299708373017,
"train_child_cross_entropy": 0.08995897322893143,
"train_parent_alignment": 0.7072291970252991,
"train_child_alignment": 0.784734480910831,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7692307692307693,
"validation/tool/silhouette_score": 0.06840348988771439,
"validation/tool/global_near_overlap_rate": 0.3269230769230769,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.054581109434366226,
"validation/parent/global_near_overlap_rate": 0.19230769230769232,
"validation/parent/top_pair/label_a": "Information Retrieval",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.13333334028720856,
"validation/parent/top_pair/centroid_cosine": 0.22527211904525757
},
{
"epoch": 16,
"train_loss": 0.24080454806486765,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 1.0,
"val_tool_retrieval_accuracy": 0.75,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.7307692307692307,
"val_parent_classification_accuracy": 0.8846153846153846,
"learning_rate": 7.980295566502464e-06,
"epoch_duration_seconds": 2.0077261100013857,
"best_val_tool_retrieval_accuracy": 0.7692307829856873,
"best_epoch": 12,
"train_parent_cross_entropy": 0.032339607261949115,
"train_child_cross_entropy": 0.060213643643591136,
"train_parent_alignment": 0.7098994188838534,
"train_child_alignment": 0.7726135386361016,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.75,
"validation/tool/silhouette_score": 0.07184359431266785,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.05527704581618309,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Information Retrieval",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.13333334028720856,
"validation/parent/top_pair/centroid_cosine": 0.20723675191402435
},
{
"epoch": 17,
"train_loss": 0.21917742159631518,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 1.0,
"val_tool_retrieval_accuracy": 0.7692307829856873,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.75,
"val_parent_classification_accuracy": 0.9038461538461539,
"learning_rate": 7.093596059113301e-06,
"epoch_duration_seconds": 1.297605815998395,
"best_val_tool_retrieval_accuracy": 0.7692307829856873,
"best_epoch": 12,
"train_parent_cross_entropy": 0.019533097743988037,
"train_child_cross_entropy": 0.05424659254236354,
"train_parent_alignment": 0.6934020320574442,
"train_child_alignment": 0.7605752746264139,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7692307692307693,
"validation/tool/silhouette_score": 0.0746900886297226,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.06387268751859665,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Information Retrieval",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.13333334028720856,
"validation/parent/top_pair/centroid_cosine": 0.17757676541805267
},
{
"epoch": 18,
"train_loss": 0.2239979604880015,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 1.0,
"val_tool_retrieval_accuracy": 0.7692307829856873,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.75,
"val_parent_classification_accuracy": 0.9038461538461539,
"learning_rate": 6.206896551724138e-06,
"epoch_duration_seconds": 1.3268793469978846,
"best_val_tool_retrieval_accuracy": 0.7692307829856873,
"best_epoch": 12,
"train_parent_cross_entropy": 0.021013346293734178,
"train_child_cross_entropy": 0.057985921286874346,
"train_parent_alignment": 0.6833301186561584,
"train_child_alignment": 0.766656776269277,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7692307692307693,
"validation/tool/silhouette_score": 0.07859379798173904,
"validation/tool/global_near_overlap_rate": 0.3269230769230769,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.06772759556770325,
"validation/parent/global_near_overlap_rate": 0.15384615384615385,
"validation/parent/top_pair/label_a": "Information Retrieval",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.13333334028720856,
"validation/parent/top_pair/centroid_cosine": 0.15037700533866882
},
{
"epoch": 19,
"train_loss": 0.2052950064341227,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 1.0,
"val_tool_retrieval_accuracy": 0.7884615659713745,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.7692307692307693,
"val_parent_classification_accuracy": 0.9038461538461539,
"learning_rate": 5.320197044334976e-06,
"epoch_duration_seconds": 1.2910668359982083,
"best_val_tool_retrieval_accuracy": 0.7884615659713745,
"best_epoch": 19,
"train_parent_cross_entropy": 0.01799314386314816,
"train_child_cross_entropy": 0.04309294786718157,
"train_parent_alignment": 0.6805168125364516,
"train_child_alignment": 0.7615723013877869,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7884615384615384,
"validation/tool/silhouette_score": 0.08252660185098648,
"validation/tool/global_near_overlap_rate": 0.3269230769230769,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.0690055787563324,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.10000000149011612,
"validation/parent/top_pair/centroid_cosine": 0.3109726309776306
},
{
"epoch": 20,
"train_loss": 0.1988206985923979,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 0.9961538314819336,
"val_tool_retrieval_accuracy": 0.7884615659713745,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.75,
"val_parent_classification_accuracy": 0.8846153846153846,
"learning_rate": 4.4334975369458135e-06,
"epoch_duration_seconds": 1.9614776899979915,
"best_val_tool_retrieval_accuracy": 0.7884615659713745,
"best_epoch": 19,
"train_parent_cross_entropy": 0.018124247073299356,
"train_child_cross_entropy": 0.036885734647512436,
"train_parent_alignment": 0.6738381319575839,
"train_child_alignment": 0.7642690075768365,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7884615384615384,
"validation/tool/silhouette_score": 0.08244223147630692,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.07241550087928772,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.10000000149011612,
"validation/parent/top_pair/centroid_cosine": 0.2947779893875122
},
{
"epoch": 21,
"train_loss": 0.19204935597048867,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 0.9961538314819336,
"val_tool_retrieval_accuracy": 0.7884615659713745,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.7307692307692307,
"val_parent_classification_accuracy": 0.8653846153846154,
"learning_rate": 3.5467980295566506e-06,
"epoch_duration_seconds": 1.3060714730017935,
"best_val_tool_retrieval_accuracy": 0.7884615659713745,
"best_epoch": 19,
"train_parent_cross_entropy": 0.017510208094285593,
"train_child_cross_entropy": 0.03250948815710015,
"train_parent_alignment": 0.6696353885862563,
"train_child_alignment": 0.7506611810790168,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7884615384615384,
"validation/tool/silhouette_score": 0.0829564779996872,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.07466989010572433,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.10000000149011612,
"validation/parent/top_pair/centroid_cosine": 0.28267771005630493
},
{
"epoch": 22,
"train_loss": 0.1794411970509423,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 1.0,
"val_tool_retrieval_accuracy": 0.7884615659713745,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.75,
"val_parent_classification_accuracy": 0.9038461538461539,
"learning_rate": 2.660098522167488e-06,
"epoch_duration_seconds": 1.3172019779958646,
"best_val_tool_retrieval_accuracy": 0.7884615659713745,
"best_epoch": 19,
"train_parent_cross_entropy": 0.014070622726447053,
"train_child_cross_entropy": 0.026905071611205738,
"train_parent_alignment": 0.6510205997361077,
"train_child_alignment": 0.733634372552236,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7884615384615384,
"validation/tool/silhouette_score": 0.08473511040210724,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.0743960365653038,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.10000000149011612,
"validation/parent/top_pair/centroid_cosine": 0.2789847254753113
},
{
"epoch": 23,
"train_loss": 0.18609628081321716,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 1.0,
"val_tool_retrieval_accuracy": 0.7884615659713745,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.75,
"val_parent_classification_accuracy": 0.9038461538461539,
"learning_rate": 1.7733990147783253e-06,
"epoch_duration_seconds": 1.2904535019988543,
"best_val_tool_retrieval_accuracy": 0.7884615659713745,
"best_epoch": 19,
"train_parent_cross_entropy": 0.01543032807401485,
"train_child_cross_entropy": 0.030358499122990504,
"train_parent_alignment": 0.6581042408943176,
"train_child_alignment": 0.7449702355596755,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7884615384615384,
"validation/tool/silhouette_score": 0.08613520115613937,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.07368122041225433,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.10000000149011612,
"validation/parent/top_pair/centroid_cosine": 0.277476966381073
},
{
"epoch": 24,
"train_loss": 0.1794088515970442,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 1.0,
"val_tool_retrieval_accuracy": 0.7884615659713745,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.7692307692307693,
"val_parent_classification_accuracy": 0.9230769230769231,
"learning_rate": 8.866995073891626e-07,
"epoch_duration_seconds": 1.3085527120056213,
"best_val_tool_retrieval_accuracy": 0.7884615659713745,
"best_epoch": 19,
"train_parent_cross_entropy": 0.01522777270939615,
"train_child_cross_entropy": 0.025616472793949976,
"train_parent_alignment": 0.6548879875077142,
"train_child_alignment": 0.7307580312093099,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7884615384615384,
"validation/tool/silhouette_score": 0.08699968457221985,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.07333475351333618,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.10000000149011612,
"validation/parent/top_pair/centroid_cosine": 0.2760659158229828
},
{
"epoch": 25,
"train_loss": 0.18173575732443067,
"train_tool_retrieval_accuracy": 0.9961538314819336,
"train_parent_retrieval_accuracy": 1.0,
"val_tool_retrieval_accuracy": 0.7884615659713745,
"val_parent_retrieval_accuracy": 0.9038461446762085,
"val_tool_classification_accuracy": 0.75,
"val_parent_classification_accuracy": 0.9038461538461539,
"learning_rate": 0.0,
"epoch_duration_seconds": 1.3227032190043246,
"best_val_tool_retrieval_accuracy": 0.7884615659713745,
"best_epoch": 19,
"train_parent_cross_entropy": 0.01568867434333596,
"train_child_cross_entropy": 0.027332179869214695,
"train_parent_alignment": 0.6537897917959425,
"train_child_alignment": 0.733359158039093,
"validation/tool/diagnostic_sample_size": 52,
"validation/tool/label_count": 26,
"validation/tool/subset_retrieval_accuracy": 0.7884615384615384,
"validation/tool/silhouette_score": 0.08730607479810715,
"validation/tool/global_near_overlap_rate": 0.3076923076923077,
"validation/parent/diagnostic_sample_size": 52,
"validation/parent/label_count": 6,
"validation/parent/subset_retrieval_accuracy": 0.9038461538461539,
"validation/parent/silhouette_score": 0.07345665991306305,
"validation/parent/global_near_overlap_rate": 0.17307692307692307,
"validation/parent/top_pair/label_a": "Computation",
"validation/parent/top_pair/label_b": "Memory",
"validation/parent/top_pair/mutual_overlap": 0.10000000149011612,
"validation/parent/top_pair/centroid_cosine": 0.27413174510002136
}
]