| { |
| "best_global_step": null, |
| "best_metric": null, |
| "best_model_checkpoint": null, |
| "epoch": 0.9880028228652082, |
| "eval_steps": 500, |
| "global_step": 4200, |
| "is_hyper_param_search": false, |
| "is_local_process_zero": true, |
| "is_world_process_zero": true, |
| "log_history": [ |
| { |
| "epoch": 0, |
| "eval/E1/acc_on_CONSISTENT": 0.7121951219512195, |
| "eval/E1/acc_on_INCONSISTENT": 0.7111111111111111, |
| "eval/E1/accuracy": 0.712, |
| "eval/E1/f1_INC": 0.47058823529411775, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.3516483516483517, |
| "eval/E1/recall_INC": 0.7111111111111111, |
| "eval/E2/acc_on_CONSISTENT": 0.9447513812154696, |
| "eval/E2/acc_on_INCONSISTENT": 0.7536231884057971, |
| "eval/E2/accuracy": 0.892, |
| "eval/E2/f1_INC": 0.7938931297709925, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.8387096774193549, |
| "eval/E2/recall_INC": 0.7536231884057971, |
| "eval/E3/acc_on_CONSISTENT": 0.9880952380952381, |
| "eval/E3/acc_on_INCONSISTENT": 0.13414634146341464, |
| "eval/E3/accuracy": 0.708, |
| "eval/E3/f1_INC": 0.23157894736842108, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.8461538461538461, |
| "eval/E3/recall_INC": 0.13414634146341464, |
| "eval/E4/acc_on_CONSISTENT": 1.0, |
| "eval/E4/acc_on_INCONSISTENT": 0.05263157894736842, |
| "eval/E4/accuracy": 0.64, |
| "eval/E4/f1_INC": 0.1, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 1.0, |
| "eval/E4/recall_INC": 0.05263157894736842, |
| "eval/edge_macro_accuracy": 0.738, |
| "eval/edge_macro_f1_INC": 0.3990150781083828, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.8349514563106796, |
| "eval/overall/acc_on_INCONSISTENT": 0.7278911564625851, |
| "eval/overall/accuracy": 0.772, |
| "eval/overall/f1_INC": 0.7896678966789669, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.8629032258064516, |
| "eval/overall/recall_INC": 0.7278911564625851, |
| "eval/overall_exact_match": 0.4, |
| "eval_label": "start", |
| "step": 0 |
| }, |
| { |
| "epoch": 0.004704775346977182, |
| "grad_norm": 0.5523808598518372, |
| "learning_rate": 1.484375e-05, |
| "loss": 0.13757184743881226, |
| "step": 20 |
| }, |
| { |
| "epoch": 0.009409550693954364, |
| "grad_norm": 0.1797914355993271, |
| "learning_rate": 3.0468750000000002e-05, |
| "loss": 0.04572285413742065, |
| "step": 40 |
| }, |
| { |
| "epoch": 0.014114326040931546, |
| "grad_norm": 0.20825961232185364, |
| "learning_rate": 4.609375e-05, |
| "loss": 0.03505098521709442, |
| "step": 60 |
| }, |
| { |
| "epoch": 0.018819101387908727, |
| "grad_norm": 0.10741130262613297, |
| "learning_rate": 6.171875e-05, |
| "loss": 0.04117105603218078, |
| "step": 80 |
| }, |
| { |
| "epoch": 0.023523876734885908, |
| "eval/E1/acc_on_CONSISTENT": 0.9317073170731708, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.804, |
| "eval/E1/f1_INC": 0.2898550724637681, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.4166666666666667, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.9171270718232044, |
| "eval/E2/acc_on_INCONSISTENT": 0.927536231884058, |
| "eval/E2/accuracy": 0.92, |
| "eval/E2/f1_INC": 0.8648648648648648, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.810126582278481, |
| "eval/E2/recall_INC": 0.927536231884058, |
| "eval/E3/acc_on_CONSISTENT": 0.6428571428571429, |
| "eval/E3/acc_on_INCONSISTENT": 0.8658536585365854, |
| "eval/E3/accuracy": 0.716, |
| "eval/E3/f1_INC": 0.6666666666666667, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.5419847328244275, |
| "eval/E3/recall_INC": 0.8658536585365854, |
| "eval/E4/acc_on_CONSISTENT": 0.8774193548387097, |
| "eval/E4/acc_on_INCONSISTENT": 0.7473684210526316, |
| "eval/E4/accuracy": 0.828, |
| "eval/E4/f1_INC": 0.7675675675675675, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.7888888888888889, |
| "eval/E4/recall_INC": 0.7473684210526316, |
| "eval/edge_macro_accuracy": 0.817, |
| "eval/edge_macro_f1_INC": 0.6472385428907168, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.6116504854368932, |
| "eval/overall/acc_on_INCONSISTENT": 0.8639455782312925, |
| "eval/overall/accuracy": 0.76, |
| "eval/overall/f1_INC": 0.8089171974522292, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.7604790419161677, |
| "eval/overall/recall_INC": 0.8639455782312925, |
| "eval/overall_exact_match": 0.488, |
| "eval_label": "step_100", |
| "step": 100 |
| }, |
| { |
| "epoch": 0.023523876734885908, |
| "grad_norm": 0.19869831204414368, |
| "learning_rate": 7.734375e-05, |
| "loss": 0.04022560715675354, |
| "step": 100 |
| }, |
| { |
| "epoch": 0.028228652081863093, |
| "grad_norm": 0.14735017716884613, |
| "learning_rate": 9.296875e-05, |
| "loss": 0.03491023778915405, |
| "step": 120 |
| }, |
| { |
| "epoch": 0.03293342742884027, |
| "grad_norm": 0.11853974312543869, |
| "learning_rate": 9.973320397768616e-05, |
| "loss": 0.03504326343536377, |
| "step": 140 |
| }, |
| { |
| "epoch": 0.037638202775817454, |
| "grad_norm": 0.2488207221031189, |
| "learning_rate": 9.924812030075187e-05, |
| "loss": 0.03231909275054932, |
| "step": 160 |
| }, |
| { |
| "epoch": 0.04234297812279464, |
| "grad_norm": 0.1328137069940567, |
| "learning_rate": 9.876303662381762e-05, |
| "loss": 0.03688516616821289, |
| "step": 180 |
| }, |
| { |
| "epoch": 0.047047753469771816, |
| "eval/E1/acc_on_CONSISTENT": 0.9951219512195122, |
| "eval/E1/acc_on_INCONSISTENT": 0.044444444444444446, |
| "eval/E1/accuracy": 0.824, |
| "eval/E1/f1_INC": 0.08333333333333334, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6666666666666666, |
| "eval/E1/recall_INC": 0.044444444444444446, |
| "eval/E2/acc_on_CONSISTENT": 0.9171270718232044, |
| "eval/E2/acc_on_INCONSISTENT": 1.0, |
| "eval/E2/accuracy": 0.94, |
| "eval/E2/f1_INC": 0.9019607843137255, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.8214285714285714, |
| "eval/E2/recall_INC": 1.0, |
| "eval/E3/acc_on_CONSISTENT": 0.9464285714285714, |
| "eval/E3/acc_on_INCONSISTENT": 0.7317073170731707, |
| "eval/E3/accuracy": 0.876, |
| "eval/E3/f1_INC": 0.7947019867549668, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.8695652173913043, |
| "eval/E3/recall_INC": 0.7317073170731707, |
| "eval/E4/acc_on_CONSISTENT": 0.9032258064516129, |
| "eval/E4/acc_on_INCONSISTENT": 0.7789473684210526, |
| "eval/E4/accuracy": 0.856, |
| "eval/E4/f1_INC": 0.8043478260869564, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8314606741573034, |
| "eval/E4/recall_INC": 0.7789473684210526, |
| "eval/edge_macro_accuracy": 0.874, |
| "eval/edge_macro_f1_INC": 0.6460859826222455, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.8349514563106796, |
| "eval/overall/acc_on_INCONSISTENT": 0.8639455782312925, |
| "eval/overall/accuracy": 0.852, |
| "eval/overall/f1_INC": 0.872852233676976, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.8819444444444444, |
| "eval/overall/recall_INC": 0.8639455782312925, |
| "eval/overall_exact_match": 0.612, |
| "eval_label": "step_200", |
| "step": 200 |
| }, |
| { |
| "epoch": 0.047047753469771816, |
| "grad_norm": 0.08311431109905243, |
| "learning_rate": 9.827795294688334e-05, |
| "loss": 0.0254146546125412, |
| "step": 200 |
| }, |
| { |
| "epoch": 0.051752528816749, |
| "grad_norm": 0.1718927025794983, |
| "learning_rate": 9.779286926994907e-05, |
| "loss": 0.027857154607772827, |
| "step": 220 |
| }, |
| { |
| "epoch": 0.056457304163726185, |
| "grad_norm": 0.1183571070432663, |
| "learning_rate": 9.73077855930148e-05, |
| "loss": 0.02674247622489929, |
| "step": 240 |
| }, |
| { |
| "epoch": 0.06116207951070336, |
| "grad_norm": 0.1724250167608261, |
| "learning_rate": 9.682270191608053e-05, |
| "loss": 0.02922247052192688, |
| "step": 260 |
| }, |
| { |
| "epoch": 0.06586685485768054, |
| "grad_norm": 0.10078955441713333, |
| "learning_rate": 9.633761823914626e-05, |
| "loss": 0.030029726028442384, |
| "step": 280 |
| }, |
| { |
| "epoch": 0.07057163020465773, |
| "eval/E1/acc_on_CONSISTENT": 0.9707317073170731, |
| "eval/E1/acc_on_INCONSISTENT": 0.08888888888888889, |
| "eval/E1/accuracy": 0.812, |
| "eval/E1/f1_INC": 0.14545454545454545, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.4, |
| "eval/E1/recall_INC": 0.08888888888888889, |
| "eval/E2/acc_on_CONSISTENT": 0.9613259668508287, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.956, |
| "eval/E2/f1_INC": 0.9219858156028369, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9027777777777778, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.7682926829268293, |
| "eval/E3/accuracy": 0.908, |
| "eval/E3/f1_INC": 0.8456375838926175, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9402985074626866, |
| "eval/E3/recall_INC": 0.7682926829268293, |
| "eval/E4/acc_on_CONSISTENT": 0.9548387096774194, |
| "eval/E4/acc_on_INCONSISTENT": 0.7894736842105263, |
| "eval/E4/accuracy": 0.892, |
| "eval/E4/f1_INC": 0.847457627118644, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9146341463414634, |
| "eval/E4/recall_INC": 0.7894736842105263, |
| "eval/edge_macro_accuracy": 0.892, |
| "eval/edge_macro_f1_INC": 0.690133893017161, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.970873786407767, |
| "eval/overall/acc_on_INCONSISTENT": 0.8571428571428571, |
| "eval/overall/accuracy": 0.904, |
| "eval/overall/f1_INC": 0.9130434782608695, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9767441860465116, |
| "eval/overall/recall_INC": 0.8571428571428571, |
| "eval/overall_exact_match": 0.66, |
| "eval_label": "step_300", |
| "step": 300 |
| }, |
| { |
| "epoch": 0.07057163020465773, |
| "grad_norm": 0.06914368271827698, |
| "learning_rate": 9.585253456221198e-05, |
| "loss": 0.028232046961784364, |
| "step": 300 |
| }, |
| { |
| "epoch": 0.07527640555163491, |
| "grad_norm": 0.28997278213500977, |
| "learning_rate": 9.536745088527772e-05, |
| "loss": 0.02833353579044342, |
| "step": 320 |
| }, |
| { |
| "epoch": 0.07998118089861209, |
| "grad_norm": 0.09726424515247345, |
| "learning_rate": 9.488236720834345e-05, |
| "loss": 0.030119282007217408, |
| "step": 340 |
| }, |
| { |
| "epoch": 0.08468595624558928, |
| "grad_norm": 0.1231047585606575, |
| "learning_rate": 9.439728353140918e-05, |
| "loss": 0.030776819586753844, |
| "step": 360 |
| }, |
| { |
| "epoch": 0.08939073159256646, |
| "grad_norm": 0.11107604950666428, |
| "learning_rate": 9.39121998544749e-05, |
| "loss": 0.025908437371253968, |
| "step": 380 |
| }, |
| { |
| "epoch": 0.09409550693954363, |
| "eval/E1/acc_on_CONSISTENT": 1.0, |
| "eval/E1/acc_on_INCONSISTENT": 0.08888888888888889, |
| "eval/E1/accuracy": 0.836, |
| "eval/E1/f1_INC": 0.163265306122449, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 1.0, |
| "eval/E1/recall_INC": 0.08888888888888889, |
| "eval/E2/acc_on_CONSISTENT": 0.9834254143646409, |
| "eval/E2/acc_on_INCONSISTENT": 0.9130434782608695, |
| "eval/E2/accuracy": 0.964, |
| "eval/E2/f1_INC": 0.9333333333333332, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9545454545454546, |
| "eval/E2/recall_INC": 0.9130434782608695, |
| "eval/E3/acc_on_CONSISTENT": 0.9821428571428571, |
| "eval/E3/acc_on_INCONSISTENT": 0.7926829268292683, |
| "eval/E3/accuracy": 0.92, |
| "eval/E3/f1_INC": 0.8666666666666668, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9558823529411765, |
| "eval/E3/recall_INC": 0.7926829268292683, |
| "eval/E4/acc_on_CONSISTENT": 0.9548387096774194, |
| "eval/E4/acc_on_INCONSISTENT": 0.8631578947368421, |
| "eval/E4/accuracy": 0.92, |
| "eval/E4/f1_INC": 0.891304347826087, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9213483146067416, |
| "eval/E4/recall_INC": 0.8631578947368421, |
| "eval/edge_macro_accuracy": 0.91, |
| "eval/edge_macro_f1_INC": 0.713642413487134, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.970873786407767, |
| "eval/overall/acc_on_INCONSISTENT": 0.8503401360544217, |
| "eval/overall/accuracy": 0.9, |
| "eval/overall/f1_INC": 0.9090909090909091, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9765625, |
| "eval/overall/recall_INC": 0.8503401360544217, |
| "eval/overall_exact_match": 0.712, |
| "eval_label": "step_400", |
| "step": 400 |
| }, |
| { |
| "epoch": 0.09409550693954363, |
| "grad_norm": 0.17487509548664093, |
| "learning_rate": 9.342711617754064e-05, |
| "loss": 0.028258711099624634, |
| "step": 400 |
| }, |
| { |
| "epoch": 0.09880028228652082, |
| "grad_norm": 0.19296877086162567, |
| "learning_rate": 9.294203250060635e-05, |
| "loss": 0.026733744144439697, |
| "step": 420 |
| }, |
| { |
| "epoch": 0.103505057633498, |
| "grad_norm": 0.14468561112880707, |
| "learning_rate": 9.245694882367208e-05, |
| "loss": 0.027249616384506226, |
| "step": 440 |
| }, |
| { |
| "epoch": 0.10820983298047518, |
| "grad_norm": 0.1229679062962532, |
| "learning_rate": 9.197186514673781e-05, |
| "loss": 0.0247579351067543, |
| "step": 460 |
| }, |
| { |
| "epoch": 0.11291460832745237, |
| "grad_norm": 0.12781408429145813, |
| "learning_rate": 9.148678146980355e-05, |
| "loss": 0.02346014231443405, |
| "step": 480 |
| }, |
| { |
| "epoch": 0.11761938367442955, |
| "eval/E1/acc_on_CONSISTENT": 0.9365853658536586, |
| "eval/E1/acc_on_INCONSISTENT": 0.4444444444444444, |
| "eval/E1/accuracy": 0.848, |
| "eval/E1/f1_INC": 0.5128205128205128, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6060606060606061, |
| "eval/E1/recall_INC": 0.4444444444444444, |
| "eval/E2/acc_on_CONSISTENT": 0.994475138121547, |
| "eval/E2/acc_on_INCONSISTENT": 0.8840579710144928, |
| "eval/E2/accuracy": 0.964, |
| "eval/E2/f1_INC": 0.9312977099236642, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9838709677419355, |
| "eval/E2/recall_INC": 0.8840579710144928, |
| "eval/E3/acc_on_CONSISTENT": 0.9702380952380952, |
| "eval/E3/acc_on_INCONSISTENT": 0.8292682926829268, |
| "eval/E3/accuracy": 0.924, |
| "eval/E3/f1_INC": 0.8774193548387096, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9315068493150684, |
| "eval/E3/recall_INC": 0.8292682926829268, |
| "eval/E4/acc_on_CONSISTENT": 0.9032258064516129, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.896, |
| "eval/E4/f1_INC": 0.865979381443299, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8484848484848485, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.908, |
| "eval/edge_macro_f1_INC": 0.7968792397565464, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9223300970873787, |
| "eval/overall/acc_on_INCONSISTENT": 0.8435374149659864, |
| "eval/overall/accuracy": 0.876, |
| "eval/overall/f1_INC": 0.8888888888888888, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9393939393939394, |
| "eval/overall/recall_INC": 0.8435374149659864, |
| "eval/overall_exact_match": 0.724, |
| "eval_label": "step_500", |
| "step": 500 |
| }, |
| { |
| "epoch": 0.11761938367442955, |
| "grad_norm": 0.14748971164226532, |
| "learning_rate": 9.100169779286928e-05, |
| "loss": 0.02654794752597809, |
| "step": 500 |
| }, |
| { |
| "epoch": 0.12232415902140673, |
| "grad_norm": 0.08809886127710342, |
| "learning_rate": 9.051661411593501e-05, |
| "loss": 0.02417362481355667, |
| "step": 520 |
| }, |
| { |
| "epoch": 0.12702893436838392, |
| "grad_norm": 0.20566488802433014, |
| "learning_rate": 9.003153043900074e-05, |
| "loss": 0.025752419233322145, |
| "step": 540 |
| }, |
| { |
| "epoch": 0.13173370971536108, |
| "grad_norm": 0.09848962724208832, |
| "learning_rate": 8.954644676206646e-05, |
| "loss": 0.026842114329338074, |
| "step": 560 |
| }, |
| { |
| "epoch": 0.13643848506233827, |
| "grad_norm": 0.13620898127555847, |
| "learning_rate": 8.906136308513218e-05, |
| "loss": 0.025594562292099, |
| "step": 580 |
| }, |
| { |
| "epoch": 0.14114326040931546, |
| "eval/E1/acc_on_CONSISTENT": 0.9804878048780488, |
| "eval/E1/acc_on_INCONSISTENT": 0.3111111111111111, |
| "eval/E1/accuracy": 0.86, |
| "eval/E1/f1_INC": 0.44444444444444436, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.7777777777777778, |
| "eval/E1/recall_INC": 0.3111111111111111, |
| "eval/E2/acc_on_CONSISTENT": 0.9779005524861878, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.972, |
| "eval/E2/f1_INC": 0.9496402877697843, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9428571428571428, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9285714285714286, |
| "eval/E3/acc_on_INCONSISTENT": 0.8902439024390244, |
| "eval/E3/accuracy": 0.916, |
| "eval/E3/f1_INC": 0.874251497005988, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.8588235294117647, |
| "eval/E3/recall_INC": 0.8902439024390244, |
| "eval/E4/acc_on_CONSISTENT": 0.9161290322580645, |
| "eval/E4/acc_on_INCONSISTENT": 0.8631578947368421, |
| "eval/E4/accuracy": 0.896, |
| "eval/E4/f1_INC": 0.8631578947368422, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8631578947368421, |
| "eval/E4/recall_INC": 0.8631578947368421, |
| "eval/edge_macro_accuracy": 0.911, |
| "eval/edge_macro_f1_INC": 0.7828735309892647, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.883495145631068, |
| "eval/overall/acc_on_INCONSISTENT": 0.8843537414965986, |
| "eval/overall/accuracy": 0.884, |
| "eval/overall/f1_INC": 0.8996539792387542, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9154929577464789, |
| "eval/overall/recall_INC": 0.8843537414965986, |
| "eval/overall_exact_match": 0.72, |
| "eval_label": "step_600", |
| "step": 600 |
| }, |
| { |
| "epoch": 0.14114326040931546, |
| "grad_norm": 0.060413166880607605, |
| "learning_rate": 8.857627940819791e-05, |
| "loss": 0.025246742367744445, |
| "step": 600 |
| }, |
| { |
| "epoch": 0.14584803575629263, |
| "grad_norm": 0.07322198897600174, |
| "learning_rate": 8.809119573126364e-05, |
| "loss": 0.0248195081949234, |
| "step": 620 |
| }, |
| { |
| "epoch": 0.15055281110326982, |
| "grad_norm": 0.16113685071468353, |
| "learning_rate": 8.760611205432939e-05, |
| "loss": 0.022868895530700685, |
| "step": 640 |
| }, |
| { |
| "epoch": 0.155257586450247, |
| "grad_norm": 0.11255485564470291, |
| "learning_rate": 8.712102837739511e-05, |
| "loss": 0.024284780025482178, |
| "step": 660 |
| }, |
| { |
| "epoch": 0.15996236179722417, |
| "grad_norm": 0.08370978385210037, |
| "learning_rate": 8.663594470046083e-05, |
| "loss": 0.023791204392910003, |
| "step": 680 |
| }, |
| { |
| "epoch": 0.16466713714420136, |
| "eval/E1/acc_on_CONSISTENT": 1.0, |
| "eval/E1/acc_on_INCONSISTENT": 0.1111111111111111, |
| "eval/E1/accuracy": 0.84, |
| "eval/E1/f1_INC": 0.19999999999999998, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 1.0, |
| "eval/E1/recall_INC": 0.1111111111111111, |
| "eval/E2/acc_on_CONSISTENT": 0.9723756906077348, |
| "eval/E2/acc_on_INCONSISTENT": 0.9855072463768116, |
| "eval/E2/accuracy": 0.976, |
| "eval/E2/f1_INC": 0.9577464788732394, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9315068493150684, |
| "eval/E2/recall_INC": 0.9855072463768116, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.8536585365853658, |
| "eval/E3/accuracy": 0.936, |
| "eval/E3/f1_INC": 0.8974358974358975, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9459459459459459, |
| "eval/E3/recall_INC": 0.8536585365853658, |
| "eval/E4/acc_on_CONSISTENT": 0.9225806451612903, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.908, |
| "eval/E4/f1_INC": 0.8795811518324607, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.875, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.915, |
| "eval/edge_macro_f1_INC": 0.7336908820353993, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9223300970873787, |
| "eval/overall/acc_on_INCONSISTENT": 0.8775510204081632, |
| "eval/overall/accuracy": 0.896, |
| "eval/overall/f1_INC": 0.9084507042253521, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9416058394160584, |
| "eval/overall/recall_INC": 0.8775510204081632, |
| "eval/overall_exact_match": 0.712, |
| "eval_label": "step_700", |
| "step": 700 |
| }, |
| { |
| "epoch": 0.16466713714420136, |
| "grad_norm": 0.12012548744678497, |
| "learning_rate": 8.615086102352656e-05, |
| "loss": 0.02781595289707184, |
| "step": 700 |
| }, |
| { |
| "epoch": 0.16937191249117856, |
| "grad_norm": 0.08447068184614182, |
| "learning_rate": 8.566577734659229e-05, |
| "loss": 0.024520662426948548, |
| "step": 720 |
| }, |
| { |
| "epoch": 0.17407668783815572, |
| "grad_norm": 0.06870961934328079, |
| "learning_rate": 8.518069366965802e-05, |
| "loss": 0.02158624231815338, |
| "step": 740 |
| }, |
| { |
| "epoch": 0.1787814631851329, |
| "grad_norm": 0.11423864215612411, |
| "learning_rate": 8.469560999272375e-05, |
| "loss": 0.01973051130771637, |
| "step": 760 |
| }, |
| { |
| "epoch": 0.1834862385321101, |
| "grad_norm": 0.1273324191570282, |
| "learning_rate": 8.421052631578948e-05, |
| "loss": 0.02168004512786865, |
| "step": 780 |
| }, |
| { |
| "epoch": 0.18819101387908727, |
| "eval/E1/acc_on_CONSISTENT": 0.9804878048780488, |
| "eval/E1/acc_on_INCONSISTENT": 0.3111111111111111, |
| "eval/E1/accuracy": 0.86, |
| "eval/E1/f1_INC": 0.44444444444444436, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.7777777777777778, |
| "eval/E1/recall_INC": 0.3111111111111111, |
| "eval/E2/acc_on_CONSISTENT": 0.9779005524861878, |
| "eval/E2/acc_on_INCONSISTENT": 0.9855072463768116, |
| "eval/E2/accuracy": 0.98, |
| "eval/E2/f1_INC": 0.9645390070921985, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9444444444444444, |
| "eval/E2/recall_INC": 0.9855072463768116, |
| "eval/E3/acc_on_CONSISTENT": 0.9464285714285714, |
| "eval/E3/acc_on_INCONSISTENT": 0.8658536585365854, |
| "eval/E3/accuracy": 0.92, |
| "eval/E3/f1_INC": 0.8765432098765432, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.8875, |
| "eval/E3/recall_INC": 0.8658536585365854, |
| "eval/E4/acc_on_CONSISTENT": 0.8774193548387097, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.88, |
| "eval/E4/f1_INC": 0.8484848484848484, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8155339805825242, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.91, |
| "eval/edge_macro_f1_INC": 0.7835028774745086, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.8640776699029126, |
| "eval/overall/acc_on_INCONSISTENT": 0.8775510204081632, |
| "eval/overall/accuracy": 0.872, |
| "eval/overall/f1_INC": 0.8896551724137932, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9020979020979021, |
| "eval/overall/recall_INC": 0.8775510204081632, |
| "eval/overall_exact_match": 0.696, |
| "eval_label": "step_800", |
| "step": 800 |
| }, |
| { |
| "epoch": 0.18819101387908727, |
| "grad_norm": 0.09624364227056503, |
| "learning_rate": 8.372544263885522e-05, |
| "loss": 0.022552117705345154, |
| "step": 800 |
| }, |
| { |
| "epoch": 0.19289578922606446, |
| "grad_norm": 0.07525322586297989, |
| "learning_rate": 8.324035896192093e-05, |
| "loss": 0.023621267080307005, |
| "step": 820 |
| }, |
| { |
| "epoch": 0.19760056457304165, |
| "grad_norm": 0.09463365375995636, |
| "learning_rate": 8.275527528498666e-05, |
| "loss": 0.02172262519598007, |
| "step": 840 |
| }, |
| { |
| "epoch": 0.2023053399200188, |
| "grad_norm": 0.10750160366296768, |
| "learning_rate": 8.227019160805239e-05, |
| "loss": 0.02797058820724487, |
| "step": 860 |
| }, |
| { |
| "epoch": 0.207010115266996, |
| "grad_norm": 0.1268552839756012, |
| "learning_rate": 8.178510793111812e-05, |
| "loss": 0.02256963700056076, |
| "step": 880 |
| }, |
| { |
| "epoch": 0.2117148906139732, |
| "eval/E1/acc_on_CONSISTENT": 0.9560975609756097, |
| "eval/E1/acc_on_INCONSISTENT": 0.26666666666666666, |
| "eval/E1/accuracy": 0.832, |
| "eval/E1/f1_INC": 0.36363636363636365, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.5714285714285714, |
| "eval/E1/recall_INC": 0.26666666666666666, |
| "eval/E2/acc_on_CONSISTENT": 0.9779005524861878, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.968, |
| "eval/E2/f1_INC": 0.9420289855072463, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9420289855072463, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9345238095238095, |
| "eval/E3/acc_on_INCONSISTENT": 0.9390243902439024, |
| "eval/E3/accuracy": 0.936, |
| "eval/E3/f1_INC": 0.9058823529411764, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.875, |
| "eval/E3/recall_INC": 0.9390243902439024, |
| "eval/E4/acc_on_CONSISTENT": 0.8709677419354839, |
| "eval/E4/acc_on_INCONSISTENT": 0.9368421052631579, |
| "eval/E4/accuracy": 0.896, |
| "eval/E4/f1_INC": 0.8725490196078431, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8165137614678899, |
| "eval/E4/recall_INC": 0.9368421052631579, |
| "eval/edge_macro_accuracy": 0.908, |
| "eval/edge_macro_f1_INC": 0.7710241804231573, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.8737864077669902, |
| "eval/overall/acc_on_INCONSISTENT": 0.9115646258503401, |
| "eval/overall/accuracy": 0.896, |
| "eval/overall/f1_INC": 0.9115646258503401, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9115646258503401, |
| "eval/overall/recall_INC": 0.9115646258503401, |
| "eval/overall_exact_match": 0.688, |
| "eval_label": "step_900", |
| "step": 900 |
| }, |
| { |
| "epoch": 0.2117148906139732, |
| "grad_norm": 0.06373687088489532, |
| "learning_rate": 8.130002425418385e-05, |
| "loss": 0.02163359373807907, |
| "step": 900 |
| }, |
| { |
| "epoch": 0.21641966596095036, |
| "grad_norm": 0.04184066876769066, |
| "learning_rate": 8.081494057724958e-05, |
| "loss": 0.023595696687698363, |
| "step": 920 |
| }, |
| { |
| "epoch": 0.22112444130792755, |
| "grad_norm": 0.09892212599515915, |
| "learning_rate": 8.032985690031531e-05, |
| "loss": 0.022030109167099, |
| "step": 940 |
| }, |
| { |
| "epoch": 0.22582921665490474, |
| "grad_norm": 0.04609724506735802, |
| "learning_rate": 7.984477322338104e-05, |
| "loss": 0.025409483909606935, |
| "step": 960 |
| }, |
| { |
| "epoch": 0.2305339920018819, |
| "grad_norm": 0.053641900420188904, |
| "learning_rate": 7.935968954644677e-05, |
| "loss": 0.024473431706428527, |
| "step": 980 |
| }, |
| { |
| "epoch": 0.2352387673488591, |
| "eval/E1/acc_on_CONSISTENT": 0.9658536585365853, |
| "eval/E1/acc_on_INCONSISTENT": 0.3333333333333333, |
| "eval/E1/accuracy": 0.852, |
| "eval/E1/f1_INC": 0.4477611940298507, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6818181818181818, |
| "eval/E1/recall_INC": 0.3333333333333333, |
| "eval/E2/acc_on_CONSISTENT": 0.9779005524861878, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.972, |
| "eval/E2/f1_INC": 0.9496402877697843, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9428571428571428, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9940476190476191, |
| "eval/E3/acc_on_INCONSISTENT": 0.8658536585365854, |
| "eval/E3/accuracy": 0.952, |
| "eval/E3/f1_INC": 0.9220779220779222, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9861111111111112, |
| "eval/E3/recall_INC": 0.8658536585365854, |
| "eval/E4/acc_on_CONSISTENT": 0.9419354838709677, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.92, |
| "eval/E4/f1_INC": 0.8936170212765957, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9032258064516129, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.9239999999999999, |
| "eval/edge_macro_f1_INC": 0.8032741062885382, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.941747572815534, |
| "eval/overall/acc_on_INCONSISTENT": 0.8639455782312925, |
| "eval/overall/accuracy": 0.896, |
| "eval/overall/f1_INC": 0.9071428571428571, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9548872180451128, |
| "eval/overall/recall_INC": 0.8639455782312925, |
| "eval/overall_exact_match": 0.744, |
| "eval_label": "step_1000", |
| "step": 1000 |
| }, |
| { |
| "epoch": 0.2352387673488591, |
| "grad_norm": 0.10360165685415268, |
| "learning_rate": 7.88746058695125e-05, |
| "loss": 0.023795010149478914, |
| "step": 1000 |
| }, |
| { |
| "epoch": 0.2399435426958363, |
| "grad_norm": 0.06076756864786148, |
| "learning_rate": 7.838952219257823e-05, |
| "loss": 0.022722464799880982, |
| "step": 1020 |
| }, |
| { |
| "epoch": 0.24464831804281345, |
| "grad_norm": 0.09723154455423355, |
| "learning_rate": 7.790443851564395e-05, |
| "loss": 0.021224670112133026, |
| "step": 1040 |
| }, |
| { |
| "epoch": 0.24935309338979064, |
| "grad_norm": 0.07220115512609482, |
| "learning_rate": 7.741935483870968e-05, |
| "loss": 0.01930866688489914, |
| "step": 1060 |
| }, |
| { |
| "epoch": 0.25405786873676783, |
| "grad_norm": 0.09694524109363556, |
| "learning_rate": 7.693427116177541e-05, |
| "loss": 0.025650495290756227, |
| "step": 1080 |
| }, |
| { |
| "epoch": 0.258762644083745, |
| "eval/E1/acc_on_CONSISTENT": 0.975609756097561, |
| "eval/E1/acc_on_INCONSISTENT": 0.26666666666666666, |
| "eval/E1/accuracy": 0.848, |
| "eval/E1/f1_INC": 0.3870967741935483, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.7058823529411765, |
| "eval/E1/recall_INC": 0.26666666666666666, |
| "eval/E2/acc_on_CONSISTENT": 0.9613259668508287, |
| "eval/E2/acc_on_INCONSISTENT": 0.927536231884058, |
| "eval/E2/accuracy": 0.952, |
| "eval/E2/f1_INC": 0.9142857142857144, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9014084507042254, |
| "eval/E2/recall_INC": 0.927536231884058, |
| "eval/E3/acc_on_CONSISTENT": 0.9880952380952381, |
| "eval/E3/acc_on_INCONSISTENT": 0.8780487804878049, |
| "eval/E3/accuracy": 0.952, |
| "eval/E3/f1_INC": 0.923076923076923, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.972972972972973, |
| "eval/E3/recall_INC": 0.8780487804878049, |
| "eval/E4/acc_on_CONSISTENT": 0.967741935483871, |
| "eval/E4/acc_on_INCONSISTENT": 0.8526315789473684, |
| "eval/E4/accuracy": 0.924, |
| "eval/E4/f1_INC": 0.8950276243093923, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9418604651162791, |
| "eval/E4/recall_INC": 0.8526315789473684, |
| "eval/edge_macro_accuracy": 0.919, |
| "eval/edge_macro_f1_INC": 0.7798717589663945, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9611650485436893, |
| "eval/overall/acc_on_INCONSISTENT": 0.8639455782312925, |
| "eval/overall/accuracy": 0.904, |
| "eval/overall/f1_INC": 0.9136690647482014, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9694656488549618, |
| "eval/overall/recall_INC": 0.8639455782312925, |
| "eval/overall_exact_match": 0.74, |
| "eval_label": "step_1100", |
| "step": 1100 |
| }, |
| { |
| "epoch": 0.258762644083745, |
| "grad_norm": 0.06381850689649582, |
| "learning_rate": 7.644918748484113e-05, |
| "loss": 0.02168910503387451, |
| "step": 1100 |
| }, |
| { |
| "epoch": 0.26346741943072216, |
| "grad_norm": 0.12510626018047333, |
| "learning_rate": 7.596410380790687e-05, |
| "loss": 0.02265911251306534, |
| "step": 1120 |
| }, |
| { |
| "epoch": 0.2681721947776994, |
| "grad_norm": 0.17503248155117035, |
| "learning_rate": 7.54790201309726e-05, |
| "loss": 0.022301320731639863, |
| "step": 1140 |
| }, |
| { |
| "epoch": 0.27287697012467654, |
| "grad_norm": 0.048879776149988174, |
| "learning_rate": 7.499393645403833e-05, |
| "loss": 0.02423481345176697, |
| "step": 1160 |
| }, |
| { |
| "epoch": 0.2775817454716537, |
| "grad_norm": 0.06504032015800476, |
| "learning_rate": 7.450885277710406e-05, |
| "loss": 0.02036643922328949, |
| "step": 1180 |
| }, |
| { |
| "epoch": 0.2822865208186309, |
| "eval/E1/acc_on_CONSISTENT": 0.9463414634146341, |
| "eval/E1/acc_on_INCONSISTENT": 0.4444444444444444, |
| "eval/E1/accuracy": 0.856, |
| "eval/E1/f1_INC": 0.5263157894736842, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6451612903225806, |
| "eval/E1/recall_INC": 0.4444444444444444, |
| "eval/E2/acc_on_CONSISTENT": 0.9558011049723757, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.956, |
| "eval/E2/f1_INC": 0.9230769230769231, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.8918918918918919, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9523809523809523, |
| "eval/E3/acc_on_INCONSISTENT": 0.8902439024390244, |
| "eval/E3/accuracy": 0.932, |
| "eval/E3/f1_INC": 0.8957055214723927, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9012345679012346, |
| "eval/E3/recall_INC": 0.8902439024390244, |
| "eval/E4/acc_on_CONSISTENT": 0.9096774193548387, |
| "eval/E4/acc_on_INCONSISTENT": 0.8947368421052632, |
| "eval/E4/accuracy": 0.904, |
| "eval/E4/f1_INC": 0.8762886597938143, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8585858585858586, |
| "eval/E4/recall_INC": 0.8947368421052632, |
| "eval/edge_macro_accuracy": 0.912, |
| "eval/edge_macro_f1_INC": 0.8053467234542036, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.883495145631068, |
| "eval/overall/acc_on_INCONSISTENT": 0.9047619047619048, |
| "eval/overall/accuracy": 0.896, |
| "eval/overall/f1_INC": 0.910958904109589, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9172413793103448, |
| "eval/overall/recall_INC": 0.9047619047619048, |
| "eval/overall_exact_match": 0.724, |
| "eval_label": "step_1200", |
| "step": 1200 |
| }, |
| { |
| "epoch": 0.2822865208186309, |
| "grad_norm": 0.06222112104296684, |
| "learning_rate": 7.402376910016979e-05, |
| "loss": 0.022185307741165162, |
| "step": 1200 |
| }, |
| { |
| "epoch": 0.2869912961656081, |
| "grad_norm": 0.06302153319120407, |
| "learning_rate": 7.35386854232355e-05, |
| "loss": 0.02110796421766281, |
| "step": 1220 |
| }, |
| { |
| "epoch": 0.29169607151258525, |
| "grad_norm": 0.0401674322783947, |
| "learning_rate": 7.305360174630123e-05, |
| "loss": 0.019500908255577088, |
| "step": 1240 |
| }, |
| { |
| "epoch": 0.2964008468595625, |
| "grad_norm": 0.0513957142829895, |
| "learning_rate": 7.256851806936696e-05, |
| "loss": 0.0228900283575058, |
| "step": 1260 |
| }, |
| { |
| "epoch": 0.30110562220653964, |
| "grad_norm": 0.08925433456897736, |
| "learning_rate": 7.20834343924327e-05, |
| "loss": 0.02407449185848236, |
| "step": 1280 |
| }, |
| { |
| "epoch": 0.3058103975535168, |
| "eval/E1/acc_on_CONSISTENT": 0.9902439024390244, |
| "eval/E1/acc_on_INCONSISTENT": 0.1111111111111111, |
| "eval/E1/accuracy": 0.832, |
| "eval/E1/f1_INC": 0.19230769230769232, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.7142857142857143, |
| "eval/E1/recall_INC": 0.1111111111111111, |
| "eval/E2/acc_on_CONSISTENT": 0.9723756906077348, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.964, |
| "eval/E2/f1_INC": 0.935251798561151, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9285714285714286, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9940476190476191, |
| "eval/E3/acc_on_INCONSISTENT": 0.8536585365853658, |
| "eval/E3/accuracy": 0.948, |
| "eval/E3/f1_INC": 0.9150326797385622, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9859154929577465, |
| "eval/E3/recall_INC": 0.8536585365853658, |
| "eval/E4/acc_on_CONSISTENT": 0.9548387096774194, |
| "eval/E4/acc_on_INCONSISTENT": 0.8736842105263158, |
| "eval/E4/accuracy": 0.924, |
| "eval/E4/f1_INC": 0.8972972972972973, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9222222222222223, |
| "eval/E4/recall_INC": 0.8736842105263158, |
| "eval/edge_macro_accuracy": 0.917, |
| "eval/edge_macro_f1_INC": 0.7349723669761757, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.941747572815534, |
| "eval/overall/acc_on_INCONSISTENT": 0.8571428571428571, |
| "eval/overall/accuracy": 0.892, |
| "eval/overall/f1_INC": 0.9032258064516128, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9545454545454546, |
| "eval/overall/recall_INC": 0.8571428571428571, |
| "eval/overall_exact_match": 0.72, |
| "eval_label": "step_1300", |
| "step": 1300 |
| }, |
| { |
| "epoch": 0.3058103975535168, |
| "grad_norm": 0.07507576048374176, |
| "learning_rate": 7.159835071549843e-05, |
| "loss": 0.022748589515686035, |
| "step": 1300 |
| }, |
| { |
| "epoch": 0.310515172900494, |
| "grad_norm": 0.09545419365167618, |
| "learning_rate": 7.111326703856416e-05, |
| "loss": 0.021193675696849823, |
| "step": 1320 |
| }, |
| { |
| "epoch": 0.3152199482474712, |
| "grad_norm": 0.06384872645139694, |
| "learning_rate": 7.062818336162989e-05, |
| "loss": 0.02309950292110443, |
| "step": 1340 |
| }, |
| { |
| "epoch": 0.31992472359444835, |
| "grad_norm": 0.04239324480295181, |
| "learning_rate": 7.01430996846956e-05, |
| "loss": 0.019398987293243408, |
| "step": 1360 |
| }, |
| { |
| "epoch": 0.32462949894142556, |
| "grad_norm": 0.09450068324804306, |
| "learning_rate": 6.965801600776134e-05, |
| "loss": 0.023485371470451356, |
| "step": 1380 |
| }, |
| { |
| "epoch": 0.32933427428840273, |
| "eval/E1/acc_on_CONSISTENT": 0.9902439024390244, |
| "eval/E1/acc_on_INCONSISTENT": 0.13333333333333333, |
| "eval/E1/accuracy": 0.836, |
| "eval/E1/f1_INC": 0.22641509433962267, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.75, |
| "eval/E1/recall_INC": 0.13333333333333333, |
| "eval/E2/acc_on_CONSISTENT": 0.9668508287292817, |
| "eval/E2/acc_on_INCONSISTENT": 0.9710144927536232, |
| "eval/E2/accuracy": 0.968, |
| "eval/E2/f1_INC": 0.943661971830986, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9178082191780822, |
| "eval/E2/recall_INC": 0.9710144927536232, |
| "eval/E3/acc_on_CONSISTENT": 0.9642857142857143, |
| "eval/E3/acc_on_INCONSISTENT": 0.8902439024390244, |
| "eval/E3/accuracy": 0.94, |
| "eval/E3/f1_INC": 0.9068322981366461, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9240506329113924, |
| "eval/E3/recall_INC": 0.8902439024390244, |
| "eval/E4/acc_on_CONSISTENT": 0.9096774193548387, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.9, |
| "eval/E4/f1_INC": 0.8704663212435233, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8571428571428571, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.911, |
| "eval/edge_macro_f1_INC": 0.7368439213876945, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.8737864077669902, |
| "eval/overall/acc_on_INCONSISTENT": 0.8979591836734694, |
| "eval/overall/accuracy": 0.888, |
| "eval/overall/f1_INC": 0.9041095890410958, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9103448275862069, |
| "eval/overall/recall_INC": 0.8979591836734694, |
| "eval/overall_exact_match": 0.716, |
| "eval_label": "step_1400", |
| "step": 1400 |
| }, |
| { |
| "epoch": 0.32933427428840273, |
| "grad_norm": 0.05435815826058388, |
| "learning_rate": 6.917293233082706e-05, |
| "loss": 0.020015403628349304, |
| "step": 1400 |
| }, |
| { |
| "epoch": 0.3340390496353799, |
| "grad_norm": 0.05114929750561714, |
| "learning_rate": 6.86878486538928e-05, |
| "loss": 0.018889188766479492, |
| "step": 1420 |
| }, |
| { |
| "epoch": 0.3387438249823571, |
| "grad_norm": 0.0863569900393486, |
| "learning_rate": 6.820276497695854e-05, |
| "loss": 0.022047021985054018, |
| "step": 1440 |
| }, |
| { |
| "epoch": 0.3434486003293343, |
| "grad_norm": 0.12862060964107513, |
| "learning_rate": 6.771768130002427e-05, |
| "loss": 0.023174412548542023, |
| "step": 1460 |
| }, |
| { |
| "epoch": 0.34815337567631144, |
| "grad_norm": 0.037643834948539734, |
| "learning_rate": 6.723259762308998e-05, |
| "loss": 0.022676143050193786, |
| "step": 1480 |
| }, |
| { |
| "epoch": 0.35285815102328866, |
| "eval/E1/acc_on_CONSISTENT": 0.9121951219512195, |
| "eval/E1/acc_on_INCONSISTENT": 0.5333333333333333, |
| "eval/E1/accuracy": 0.844, |
| "eval/E1/f1_INC": 0.5517241379310344, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.5714285714285714, |
| "eval/E1/recall_INC": 0.5333333333333333, |
| "eval/E2/acc_on_CONSISTENT": 0.988950276243094, |
| "eval/E2/acc_on_INCONSISTENT": 0.9710144927536232, |
| "eval/E2/accuracy": 0.984, |
| "eval/E2/f1_INC": 0.9710144927536232, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9710144927536232, |
| "eval/E2/recall_INC": 0.9710144927536232, |
| "eval/E3/acc_on_CONSISTENT": 0.9880952380952381, |
| "eval/E3/acc_on_INCONSISTENT": 0.926829268292683, |
| "eval/E3/accuracy": 0.968, |
| "eval/E3/f1_INC": 0.9500000000000001, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9743589743589743, |
| "eval/E3/recall_INC": 0.926829268292683, |
| "eval/E4/acc_on_CONSISTENT": 0.9354838709677419, |
| "eval/E4/acc_on_INCONSISTENT": 0.9052631578947369, |
| "eval/E4/accuracy": 0.924, |
| "eval/E4/f1_INC": 0.900523560209424, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8958333333333334, |
| "eval/E4/recall_INC": 0.9052631578947369, |
| "eval/edge_macro_accuracy": 0.9299999999999999, |
| "eval/edge_macro_f1_INC": 0.8433155477235204, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9320388349514563, |
| "eval/overall/acc_on_INCONSISTENT": 0.8979591836734694, |
| "eval/overall/accuracy": 0.912, |
| "eval/overall/f1_INC": 0.923076923076923, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9496402877697842, |
| "eval/overall/recall_INC": 0.8979591836734694, |
| "eval/overall_exact_match": 0.764, |
| "eval_label": "step_1500", |
| "step": 1500 |
| }, |
| { |
| "epoch": 0.35285815102328866, |
| "grad_norm": 0.06953489035367966, |
| "learning_rate": 6.674751394615571e-05, |
| "loss": 0.022417673468589784, |
| "step": 1500 |
| }, |
| { |
| "epoch": 0.3575629263702658, |
| "grad_norm": 0.11855115741491318, |
| "learning_rate": 6.626243026922144e-05, |
| "loss": 0.019565111398696898, |
| "step": 1520 |
| }, |
| { |
| "epoch": 0.362267701717243, |
| "grad_norm": 0.04791810363531113, |
| "learning_rate": 6.577734659228717e-05, |
| "loss": 0.019627970457077027, |
| "step": 1540 |
| }, |
| { |
| "epoch": 0.3669724770642202, |
| "grad_norm": 0.05584453046321869, |
| "learning_rate": 6.52922629153529e-05, |
| "loss": 0.019779224693775178, |
| "step": 1560 |
| }, |
| { |
| "epoch": 0.37167725241119737, |
| "grad_norm": 0.09220653772354126, |
| "learning_rate": 6.480717923841863e-05, |
| "loss": 0.021753139793872833, |
| "step": 1580 |
| }, |
| { |
| "epoch": 0.37638202775817453, |
| "eval/E1/acc_on_CONSISTENT": 0.9951219512195122, |
| "eval/E1/acc_on_INCONSISTENT": 0.15555555555555556, |
| "eval/E1/accuracy": 0.844, |
| "eval/E1/f1_INC": 0.26415094339622647, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.875, |
| "eval/E1/recall_INC": 0.15555555555555556, |
| "eval/E2/acc_on_CONSISTENT": 1.0, |
| "eval/E2/acc_on_INCONSISTENT": 0.8695652173913043, |
| "eval/E2/accuracy": 0.964, |
| "eval/E2/f1_INC": 0.9302325581395349, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 1.0, |
| "eval/E2/recall_INC": 0.8695652173913043, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.8536585365853658, |
| "eval/E3/accuracy": 0.936, |
| "eval/E3/f1_INC": 0.8974358974358975, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9459459459459459, |
| "eval/E3/recall_INC": 0.8536585365853658, |
| "eval/E4/acc_on_CONSISTENT": 0.9225806451612903, |
| "eval/E4/acc_on_INCONSISTENT": 0.8736842105263158, |
| "eval/E4/accuracy": 0.904, |
| "eval/E4/f1_INC": 0.8736842105263158, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8736842105263158, |
| "eval/E4/recall_INC": 0.8736842105263158, |
| "eval/edge_macro_accuracy": 0.912, |
| "eval/edge_macro_f1_INC": 0.7413759023744937, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9029126213592233, |
| "eval/overall/acc_on_INCONSISTENT": 0.8435374149659864, |
| "eval/overall/accuracy": 0.868, |
| "eval/overall/f1_INC": 0.8825622775800711, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9253731343283582, |
| "eval/overall/recall_INC": 0.8435374149659864, |
| "eval/overall_exact_match": 0.72, |
| "eval_label": "step_1600", |
| "step": 1600 |
| }, |
| { |
| "epoch": 0.37638202775817453, |
| "grad_norm": 0.0731038823723793, |
| "learning_rate": 6.432209556148437e-05, |
| "loss": 0.02192704677581787, |
| "step": 1600 |
| }, |
| { |
| "epoch": 0.38108680310515175, |
| "grad_norm": 0.04546733573079109, |
| "learning_rate": 6.383701188455009e-05, |
| "loss": 0.023643997311592103, |
| "step": 1620 |
| }, |
| { |
| "epoch": 0.3857915784521289, |
| "grad_norm": 0.11369301378726959, |
| "learning_rate": 6.335192820761581e-05, |
| "loss": 0.018663662672042846, |
| "step": 1640 |
| }, |
| { |
| "epoch": 0.3904963537991061, |
| "grad_norm": 0.12996052205562592, |
| "learning_rate": 6.286684453068154e-05, |
| "loss": 0.024522665143013, |
| "step": 1660 |
| }, |
| { |
| "epoch": 0.3952011291460833, |
| "grad_norm": 0.03448007255792618, |
| "learning_rate": 6.238176085374727e-05, |
| "loss": 0.01910881996154785, |
| "step": 1680 |
| }, |
| { |
| "epoch": 0.39990590449306046, |
| "eval/E1/acc_on_CONSISTENT": 0.975609756097561, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.84, |
| "eval/E1/f1_INC": 0.3333333333333333, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6666666666666666, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.9723756906077348, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.964, |
| "eval/E2/f1_INC": 0.935251798561151, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9285714285714286, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.8902439024390244, |
| "eval/E3/accuracy": 0.948, |
| "eval/E3/f1_INC": 0.9182389937106918, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.948051948051948, |
| "eval/E3/recall_INC": 0.8902439024390244, |
| "eval/E4/acc_on_CONSISTENT": 0.9548387096774194, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.928, |
| "eval/E4/f1_INC": 0.9032258064516129, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9230769230769231, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.9199999999999999, |
| "eval/edge_macro_f1_INC": 0.7725124830141973, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.941747572815534, |
| "eval/overall/acc_on_INCONSISTENT": 0.8707482993197279, |
| "eval/overall/accuracy": 0.9, |
| "eval/overall/f1_INC": 0.9110320284697508, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9552238805970149, |
| "eval/overall/recall_INC": 0.8707482993197279, |
| "eval/overall_exact_match": 0.728, |
| "eval_label": "step_1700", |
| "step": 1700 |
| }, |
| { |
| "epoch": 0.39990590449306046, |
| "grad_norm": 0.0586729422211647, |
| "learning_rate": 6.1896677176813e-05, |
| "loss": 0.019758285582065584, |
| "step": 1700 |
| }, |
| { |
| "epoch": 0.4046106798400376, |
| "grad_norm": 0.10014427453279495, |
| "learning_rate": 6.141159349987873e-05, |
| "loss": 0.017364802956581115, |
| "step": 1720 |
| }, |
| { |
| "epoch": 0.40931545518701484, |
| "grad_norm": 0.0491943284869194, |
| "learning_rate": 6.092650982294445e-05, |
| "loss": 0.019269563257694244, |
| "step": 1740 |
| }, |
| { |
| "epoch": 0.414020230533992, |
| "grad_norm": 0.11527211219072342, |
| "learning_rate": 6.0441426146010196e-05, |
| "loss": 0.019721153378486633, |
| "step": 1760 |
| }, |
| { |
| "epoch": 0.41872500588096917, |
| "grad_norm": 0.08479166775941849, |
| "learning_rate": 5.995634246907592e-05, |
| "loss": 0.019190581142902376, |
| "step": 1780 |
| }, |
| { |
| "epoch": 0.4234297812279464, |
| "eval/E1/acc_on_CONSISTENT": 0.9609756097560975, |
| "eval/E1/acc_on_INCONSISTENT": 0.24444444444444444, |
| "eval/E1/accuracy": 0.832, |
| "eval/E1/f1_INC": 0.34375, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.5789473684210527, |
| "eval/E1/recall_INC": 0.24444444444444444, |
| "eval/E2/acc_on_CONSISTENT": 1.0, |
| "eval/E2/acc_on_INCONSISTENT": 0.9710144927536232, |
| "eval/E2/accuracy": 0.992, |
| "eval/E2/f1_INC": 0.9852941176470589, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 1.0, |
| "eval/E2/recall_INC": 0.9710144927536232, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.8780487804878049, |
| "eval/E3/accuracy": 0.944, |
| "eval/E3/f1_INC": 0.9113924050632912, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9473684210526315, |
| "eval/E3/recall_INC": 0.8780487804878049, |
| "eval/E4/acc_on_CONSISTENT": 0.9806451612903225, |
| "eval/E4/acc_on_INCONSISTENT": 0.8631578947368421, |
| "eval/E4/accuracy": 0.936, |
| "eval/E4/f1_INC": 0.9111111111111111, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9647058823529412, |
| "eval/E4/recall_INC": 0.8631578947368421, |
| "eval/edge_macro_accuracy": 0.9259999999999999, |
| "eval/edge_macro_f1_INC": 0.7878869084553652, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9611650485436893, |
| "eval/overall/acc_on_INCONSISTENT": 0.8843537414965986, |
| "eval/overall/accuracy": 0.916, |
| "eval/overall/f1_INC": 0.9252669039145908, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9701492537313433, |
| "eval/overall/recall_INC": 0.8843537414965986, |
| "eval/overall_exact_match": 0.74, |
| "eval_label": "step_1800", |
| "step": 1800 |
| }, |
| { |
| "epoch": 0.4234297812279464, |
| "grad_norm": 0.04938641935586929, |
| "learning_rate": 5.947125879214165e-05, |
| "loss": 0.021737626194953917, |
| "step": 1800 |
| }, |
| { |
| "epoch": 0.42813455657492355, |
| "grad_norm": 0.1100970059633255, |
| "learning_rate": 5.8986175115207376e-05, |
| "loss": 0.01796163022518158, |
| "step": 1820 |
| }, |
| { |
| "epoch": 0.4328393319219007, |
| "grad_norm": 0.08918889611959457, |
| "learning_rate": 5.8501091438273106e-05, |
| "loss": 0.023652970790863037, |
| "step": 1840 |
| }, |
| { |
| "epoch": 0.43754410726887794, |
| "grad_norm": 0.16495102643966675, |
| "learning_rate": 5.801600776133883e-05, |
| "loss": 0.021101899445056915, |
| "step": 1860 |
| }, |
| { |
| "epoch": 0.4422488826158551, |
| "grad_norm": 0.12315504252910614, |
| "learning_rate": 5.753092408440456e-05, |
| "loss": 0.022888024151325227, |
| "step": 1880 |
| }, |
| { |
| "epoch": 0.44695365796283226, |
| "eval/E1/acc_on_CONSISTENT": 0.9707317073170731, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.836, |
| "eval/E1/f1_INC": 0.3278688524590164, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.625, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.9779005524861878, |
| "eval/E2/acc_on_INCONSISTENT": 1.0, |
| "eval/E2/accuracy": 0.984, |
| "eval/E2/f1_INC": 0.971830985915493, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9452054794520548, |
| "eval/E2/recall_INC": 1.0, |
| "eval/E3/acc_on_CONSISTENT": 0.9702380952380952, |
| "eval/E3/acc_on_INCONSISTENT": 0.8780487804878049, |
| "eval/E3/accuracy": 0.94, |
| "eval/E3/f1_INC": 0.9056603773584906, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.935064935064935, |
| "eval/E3/recall_INC": 0.8780487804878049, |
| "eval/E4/acc_on_CONSISTENT": 0.9870967741935484, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.948, |
| "eval/E4/f1_INC": 0.9281767955801105, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9767441860465116, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.9269999999999999, |
| "eval/edge_macro_f1_INC": 0.7833842528282776, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9514563106796117, |
| "eval/overall/acc_on_INCONSISTENT": 0.8775510204081632, |
| "eval/overall/accuracy": 0.908, |
| "eval/overall/f1_INC": 0.9181494661921707, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9626865671641791, |
| "eval/overall/recall_INC": 0.8775510204081632, |
| "eval/overall_exact_match": 0.752, |
| "eval_label": "step_1900", |
| "step": 1900 |
| }, |
| { |
| "epoch": 0.44695365796283226, |
| "grad_norm": 0.09929759055376053, |
| "learning_rate": 5.70458404074703e-05, |
| "loss": 0.025352621078491212, |
| "step": 1900 |
| }, |
| { |
| "epoch": 0.4516584333098095, |
| "grad_norm": 0.06084238737821579, |
| "learning_rate": 5.656075673053602e-05, |
| "loss": 0.01936248689889908, |
| "step": 1920 |
| }, |
| { |
| "epoch": 0.45636320865678665, |
| "grad_norm": 0.043425630778074265, |
| "learning_rate": 5.607567305360175e-05, |
| "loss": 0.017471833527088164, |
| "step": 1940 |
| }, |
| { |
| "epoch": 0.4610679840037638, |
| "grad_norm": 0.09588296711444855, |
| "learning_rate": 5.559058937666748e-05, |
| "loss": 0.017113274335861205, |
| "step": 1960 |
| }, |
| { |
| "epoch": 0.46577275935074103, |
| "grad_norm": 0.06289476901292801, |
| "learning_rate": 5.510550569973321e-05, |
| "loss": 0.0180003821849823, |
| "step": 1980 |
| }, |
| { |
| "epoch": 0.4704775346977182, |
| "eval/E1/acc_on_CONSISTENT": 0.9463414634146341, |
| "eval/E1/acc_on_INCONSISTENT": 0.35555555555555557, |
| "eval/E1/accuracy": 0.84, |
| "eval/E1/f1_INC": 0.4444444444444444, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.5925925925925926, |
| "eval/E1/recall_INC": 0.35555555555555557, |
| "eval/E2/acc_on_CONSISTENT": 0.9723756906077348, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.964, |
| "eval/E2/f1_INC": 0.935251798561151, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9285714285714286, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9464285714285714, |
| "eval/E3/acc_on_INCONSISTENT": 0.9146341463414634, |
| "eval/E3/accuracy": 0.936, |
| "eval/E3/f1_INC": 0.9036144578313253, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.8928571428571429, |
| "eval/E3/recall_INC": 0.9146341463414634, |
| "eval/E4/acc_on_CONSISTENT": 0.8903225806451613, |
| "eval/E4/acc_on_INCONSISTENT": 0.9052631578947369, |
| "eval/E4/accuracy": 0.896, |
| "eval/E4/f1_INC": 0.8686868686868687, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8349514563106796, |
| "eval/E4/recall_INC": 0.9052631578947369, |
| "eval/edge_macro_accuracy": 0.909, |
| "eval/edge_macro_f1_INC": 0.7879993923809474, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.883495145631068, |
| "eval/overall/acc_on_INCONSISTENT": 0.9183673469387755, |
| "eval/overall/accuracy": 0.904, |
| "eval/overall/f1_INC": 0.9183673469387755, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9183673469387755, |
| "eval/overall/recall_INC": 0.9183673469387755, |
| "eval/overall_exact_match": 0.716, |
| "eval_label": "step_2000", |
| "step": 2000 |
| }, |
| { |
| "epoch": 0.4704775346977182, |
| "grad_norm": 0.0629287138581276, |
| "learning_rate": 5.462042202279893e-05, |
| "loss": 0.019795671105384827, |
| "step": 2000 |
| }, |
| { |
| "epoch": 0.47518231004469536, |
| "grad_norm": 0.08561868220567703, |
| "learning_rate": 5.413533834586466e-05, |
| "loss": 0.020380768179893493, |
| "step": 2020 |
| }, |
| { |
| "epoch": 0.4798870853916726, |
| "grad_norm": 0.06038222834467888, |
| "learning_rate": 5.365025466893039e-05, |
| "loss": 0.02099545896053314, |
| "step": 2040 |
| }, |
| { |
| "epoch": 0.48459186073864974, |
| "grad_norm": 0.1428939700126648, |
| "learning_rate": 5.3165170991996126e-05, |
| "loss": 0.017899855971336365, |
| "step": 2060 |
| }, |
| { |
| "epoch": 0.4892966360856269, |
| "grad_norm": 0.06938181817531586, |
| "learning_rate": 5.2680087315061855e-05, |
| "loss": 0.02358972132205963, |
| "step": 2080 |
| }, |
| { |
| "epoch": 0.4940014114326041, |
| "eval/E1/acc_on_CONSISTENT": 0.9853658536585366, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.848, |
| "eval/E1/f1_INC": 0.3448275862068966, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.7692307692307693, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.994475138121547, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.98, |
| "eval/E2/f1_INC": 0.962962962962963, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9848484848484849, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9821428571428571, |
| "eval/E3/acc_on_INCONSISTENT": 0.9024390243902439, |
| "eval/E3/accuracy": 0.956, |
| "eval/E3/f1_INC": 0.9308176100628931, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.961038961038961, |
| "eval/E3/recall_INC": 0.9024390243902439, |
| "eval/E4/acc_on_CONSISTENT": 0.9419354838709677, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.92, |
| "eval/E4/f1_INC": 0.8936170212765957, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9032258064516129, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.9259999999999999, |
| "eval/edge_macro_f1_INC": 0.7830562951273371, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9514563106796117, |
| "eval/overall/acc_on_INCONSISTENT": 0.891156462585034, |
| "eval/overall/accuracy": 0.916, |
| "eval/overall/f1_INC": 0.9257950530035336, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9632352941176471, |
| "eval/overall/recall_INC": 0.891156462585034, |
| "eval/overall_exact_match": 0.752, |
| "eval_label": "step_2100", |
| "step": 2100 |
| }, |
| { |
| "epoch": 0.4940014114326041, |
| "grad_norm": 0.08681521564722061, |
| "learning_rate": 5.2195003638127584e-05, |
| "loss": 0.020199275016784667, |
| "step": 2100 |
| }, |
| { |
| "epoch": 0.4987061867795813, |
| "grad_norm": 0.07912258803844452, |
| "learning_rate": 5.1709919961193307e-05, |
| "loss": 0.017395436763763428, |
| "step": 2120 |
| }, |
| { |
| "epoch": 0.5034109621265584, |
| "grad_norm": 0.0859365239739418, |
| "learning_rate": 5.1224836284259036e-05, |
| "loss": 0.01904984563589096, |
| "step": 2140 |
| }, |
| { |
| "epoch": 0.5081157374735357, |
| "grad_norm": 0.061891067773103714, |
| "learning_rate": 5.0739752607324765e-05, |
| "loss": 0.020740336179733275, |
| "step": 2160 |
| }, |
| { |
| "epoch": 0.5128205128205128, |
| "grad_norm": 0.03536997735500336, |
| "learning_rate": 5.0254668930390494e-05, |
| "loss": 0.016544927656650544, |
| "step": 2180 |
| }, |
| { |
| "epoch": 0.51752528816749, |
| "eval/E1/acc_on_CONSISTENT": 0.9414634146341463, |
| "eval/E1/acc_on_INCONSISTENT": 0.4222222222222222, |
| "eval/E1/accuracy": 0.848, |
| "eval/E1/f1_INC": 0.5, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6129032258064516, |
| "eval/E1/recall_INC": 0.4222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.9723756906077348, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.964, |
| "eval/E2/f1_INC": 0.935251798561151, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9285714285714286, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.8658536585365854, |
| "eval/E3/accuracy": 0.94, |
| "eval/E3/f1_INC": 0.9044585987261147, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9466666666666667, |
| "eval/E3/recall_INC": 0.8658536585365854, |
| "eval/E4/acc_on_CONSISTENT": 0.967741935483871, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.936, |
| "eval/E4/f1_INC": 0.9130434782608695, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9438202247191011, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.9219999999999999, |
| "eval/edge_macro_f1_INC": 0.8131884688870338, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.912621359223301, |
| "eval/overall/acc_on_INCONSISTENT": 0.8843537414965986, |
| "eval/overall/accuracy": 0.896, |
| "eval/overall/f1_INC": 0.909090909090909, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.935251798561151, |
| "eval/overall/recall_INC": 0.8843537414965986, |
| "eval/overall_exact_match": 0.736, |
| "eval_label": "step_2200", |
| "step": 2200 |
| }, |
| { |
| "epoch": 0.51752528816749, |
| "grad_norm": 0.10411830991506577, |
| "learning_rate": 4.976958525345622e-05, |
| "loss": 0.022702106833457948, |
| "step": 2200 |
| }, |
| { |
| "epoch": 0.5222300635144672, |
| "grad_norm": 0.09459930658340454, |
| "learning_rate": 4.928450157652195e-05, |
| "loss": 0.01726412624120712, |
| "step": 2220 |
| }, |
| { |
| "epoch": 0.5269348388614443, |
| "grad_norm": 0.08819395303726196, |
| "learning_rate": 4.879941789958768e-05, |
| "loss": 0.02211476266384125, |
| "step": 2240 |
| }, |
| { |
| "epoch": 0.5316396142084215, |
| "grad_norm": 0.04895342141389847, |
| "learning_rate": 4.831433422265341e-05, |
| "loss": 0.01894928514957428, |
| "step": 2260 |
| }, |
| { |
| "epoch": 0.5363443895553988, |
| "grad_norm": 0.03727812319993973, |
| "learning_rate": 4.782925054571914e-05, |
| "loss": 0.024250733852386474, |
| "step": 2280 |
| }, |
| { |
| "epoch": 0.5410491649023759, |
| "eval/E1/acc_on_CONSISTENT": 0.9902439024390244, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.852, |
| "eval/E1/f1_INC": 0.3508771929824561, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.8333333333333334, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.9834254143646409, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.976, |
| "eval/E2/f1_INC": 0.9565217391304348, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9565217391304348, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9940476190476191, |
| "eval/E3/acc_on_INCONSISTENT": 0.8414634146341463, |
| "eval/E3/accuracy": 0.944, |
| "eval/E3/f1_INC": 0.9078947368421053, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9857142857142858, |
| "eval/E3/recall_INC": 0.8414634146341463, |
| "eval/E4/acc_on_CONSISTENT": 0.9548387096774194, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.928, |
| "eval/E4/f1_INC": 0.9032258064516129, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9230769230769231, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.925, |
| "eval/edge_macro_f1_INC": 0.7796298688516523, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9611650485436893, |
| "eval/overall/acc_on_INCONSISTENT": 0.8843537414965986, |
| "eval/overall/accuracy": 0.916, |
| "eval/overall/f1_INC": 0.9252669039145908, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9701492537313433, |
| "eval/overall/recall_INC": 0.8843537414965986, |
| "eval/overall_exact_match": 0.756, |
| "eval_label": "step_2300", |
| "step": 2300 |
| }, |
| { |
| "epoch": 0.5410491649023759, |
| "grad_norm": 0.052870478481054306, |
| "learning_rate": 4.734416686878487e-05, |
| "loss": 0.01879228502511978, |
| "step": 2300 |
| }, |
| { |
| "epoch": 0.5457539402493531, |
| "grad_norm": 0.03650202229619026, |
| "learning_rate": 4.68590831918506e-05, |
| "loss": 0.02177768349647522, |
| "step": 2320 |
| }, |
| { |
| "epoch": 0.5504587155963303, |
| "grad_norm": 0.0578182190656662, |
| "learning_rate": 4.637399951491633e-05, |
| "loss": 0.02012043297290802, |
| "step": 2340 |
| }, |
| { |
| "epoch": 0.5551634909433074, |
| "grad_norm": 0.09132088720798492, |
| "learning_rate": 4.5888915837982056e-05, |
| "loss": 0.02349788397550583, |
| "step": 2360 |
| }, |
| { |
| "epoch": 0.5598682662902846, |
| "grad_norm": 0.09232542663812637, |
| "learning_rate": 4.5403832161047785e-05, |
| "loss": 0.023439812660217284, |
| "step": 2380 |
| }, |
| { |
| "epoch": 0.5645730416372619, |
| "eval/E1/acc_on_CONSISTENT": 0.975609756097561, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.84, |
| "eval/E1/f1_INC": 0.3333333333333333, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6666666666666666, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.9613259668508287, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.956, |
| "eval/E2/f1_INC": 0.9219858156028369, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9027777777777778, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9583333333333334, |
| "eval/E3/acc_on_INCONSISTENT": 0.8292682926829268, |
| "eval/E3/accuracy": 0.916, |
| "eval/E3/f1_INC": 0.8662420382165605, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9066666666666666, |
| "eval/E3/recall_INC": 0.8292682926829268, |
| "eval/E4/acc_on_CONSISTENT": 0.9354838709677419, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.916, |
| "eval/E4/f1_INC": 0.8888888888888888, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8936170212765957, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.907, |
| "eval/edge_macro_f1_INC": 0.7526125190104049, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9029126213592233, |
| "eval/overall/acc_on_INCONSISTENT": 0.8639455782312925, |
| "eval/overall/accuracy": 0.88, |
| "eval/overall/f1_INC": 0.8943661971830985, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.927007299270073, |
| "eval/overall/recall_INC": 0.8639455782312925, |
| "eval/overall_exact_match": 0.716, |
| "eval_label": "step_2400", |
| "step": 2400 |
| }, |
| { |
| "epoch": 0.5645730416372619, |
| "grad_norm": 0.09791548550128937, |
| "learning_rate": 4.491874848411351e-05, |
| "loss": 0.01996140480041504, |
| "step": 2400 |
| }, |
| { |
| "epoch": 0.569277816984239, |
| "grad_norm": 0.043326977640390396, |
| "learning_rate": 4.4433664807179243e-05, |
| "loss": 0.02120434641838074, |
| "step": 2420 |
| }, |
| { |
| "epoch": 0.5739825923312162, |
| "grad_norm": 0.037307336926460266, |
| "learning_rate": 4.394858113024497e-05, |
| "loss": 0.01826932728290558, |
| "step": 2440 |
| }, |
| { |
| "epoch": 0.5786873676781934, |
| "grad_norm": 0.04846954718232155, |
| "learning_rate": 4.3463497453310695e-05, |
| "loss": 0.02175077944993973, |
| "step": 2460 |
| }, |
| { |
| "epoch": 0.5833921430251705, |
| "grad_norm": 0.10085655003786087, |
| "learning_rate": 4.2978413776376424e-05, |
| "loss": 0.020642833411693574, |
| "step": 2480 |
| }, |
| { |
| "epoch": 0.5880969183721477, |
| "eval/E1/acc_on_CONSISTENT": 0.9804878048780488, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.844, |
| "eval/E1/f1_INC": 0.3389830508474576, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.7142857142857143, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.988950276243094, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.976, |
| "eval/E2/f1_INC": 0.9558823529411764, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9701492537313433, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.7682926829268293, |
| "eval/E3/accuracy": 0.908, |
| "eval/E3/f1_INC": 0.8456375838926175, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9402985074626866, |
| "eval/E3/recall_INC": 0.7682926829268293, |
| "eval/E4/acc_on_CONSISTENT": 0.9741935483870968, |
| "eval/E4/acc_on_INCONSISTENT": 0.8421052631578947, |
| "eval/E4/accuracy": 0.924, |
| "eval/E4/f1_INC": 0.8938547486033519, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9523809523809523, |
| "eval/E4/recall_INC": 0.8421052631578947, |
| "eval/edge_macro_accuracy": 0.913, |
| "eval/edge_macro_f1_INC": 0.7585894340711509, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9514563106796117, |
| "eval/overall/acc_on_INCONSISTENT": 0.8299319727891157, |
| "eval/overall/accuracy": 0.88, |
| "eval/overall/f1_INC": 0.8905109489051095, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9606299212598425, |
| "eval/overall/recall_INC": 0.8299319727891157, |
| "eval/overall_exact_match": 0.724, |
| "eval_label": "step_2500", |
| "step": 2500 |
| }, |
| { |
| "epoch": 0.5880969183721477, |
| "grad_norm": 0.06443936377763748, |
| "learning_rate": 4.249333009944216e-05, |
| "loss": 0.017859846353530884, |
| "step": 2500 |
| }, |
| { |
| "epoch": 0.592801693719125, |
| "grad_norm": 0.11238661408424377, |
| "learning_rate": 4.200824642250788e-05, |
| "loss": 0.015081197023391724, |
| "step": 2520 |
| }, |
| { |
| "epoch": 0.597506469066102, |
| "grad_norm": 0.06273433566093445, |
| "learning_rate": 4.152316274557361e-05, |
| "loss": 0.021456241607666016, |
| "step": 2540 |
| }, |
| { |
| "epoch": 0.6022112444130793, |
| "grad_norm": 0.05472638085484505, |
| "learning_rate": 4.103807906863934e-05, |
| "loss": 0.019296465814113616, |
| "step": 2560 |
| }, |
| { |
| "epoch": 0.6069160197600565, |
| "grad_norm": 0.09310346096754074, |
| "learning_rate": 4.0552995391705076e-05, |
| "loss": 0.021483346819877625, |
| "step": 2580 |
| }, |
| { |
| "epoch": 0.6116207951070336, |
| "eval/E1/acc_on_CONSISTENT": 0.9463414634146341, |
| "eval/E1/acc_on_INCONSISTENT": 0.4888888888888889, |
| "eval/E1/accuracy": 0.864, |
| "eval/E1/f1_INC": 0.5641025641025641, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6666666666666666, |
| "eval/E1/recall_INC": 0.4888888888888889, |
| "eval/E2/acc_on_CONSISTENT": 0.9723756906077348, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.968, |
| "eval/E2/f1_INC": 0.9428571428571428, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9295774647887324, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9642857142857143, |
| "eval/E3/acc_on_INCONSISTENT": 0.8780487804878049, |
| "eval/E3/accuracy": 0.936, |
| "eval/E3/f1_INC": 0.9, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9230769230769231, |
| "eval/E3/recall_INC": 0.8780487804878049, |
| "eval/E4/acc_on_CONSISTENT": 0.9161290322580645, |
| "eval/E4/acc_on_INCONSISTENT": 0.9052631578947369, |
| "eval/E4/accuracy": 0.912, |
| "eval/E4/f1_INC": 0.88659793814433, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8686868686868687, |
| "eval/E4/recall_INC": 0.9052631578947369, |
| "eval/edge_macro_accuracy": 0.92, |
| "eval/edge_macro_f1_INC": 0.8233894112760092, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9029126213592233, |
| "eval/overall/acc_on_INCONSISTENT": 0.9047619047619048, |
| "eval/overall/accuracy": 0.904, |
| "eval/overall/f1_INC": 0.9172413793103449, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9300699300699301, |
| "eval/overall/recall_INC": 0.9047619047619048, |
| "eval/overall_exact_match": 0.752, |
| "eval_label": "step_2600", |
| "step": 2600 |
| }, |
| { |
| "epoch": 0.6116207951070336, |
| "grad_norm": 0.06032578647136688, |
| "learning_rate": 4.00679117147708e-05, |
| "loss": 0.020984357595443724, |
| "step": 2600 |
| }, |
| { |
| "epoch": 0.6163255704540108, |
| "grad_norm": 0.08005069941282272, |
| "learning_rate": 3.958282803783653e-05, |
| "loss": 0.018450188636779784, |
| "step": 2620 |
| }, |
| { |
| "epoch": 0.621030345800988, |
| "grad_norm": 0.03903232887387276, |
| "learning_rate": 3.909774436090226e-05, |
| "loss": 0.019824470579624175, |
| "step": 2640 |
| }, |
| { |
| "epoch": 0.6257351211479651, |
| "grad_norm": 0.0628625676035881, |
| "learning_rate": 3.8612660683967986e-05, |
| "loss": 0.020582506060600282, |
| "step": 2660 |
| }, |
| { |
| "epoch": 0.6304398964949424, |
| "grad_norm": 0.04103922098875046, |
| "learning_rate": 3.8127577007033715e-05, |
| "loss": 0.017628610134124756, |
| "step": 2680 |
| }, |
| { |
| "epoch": 0.6351446718419196, |
| "eval/E1/acc_on_CONSISTENT": 0.9560975609756097, |
| "eval/E1/acc_on_INCONSISTENT": 0.3111111111111111, |
| "eval/E1/accuracy": 0.84, |
| "eval/E1/f1_INC": 0.411764705882353, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6086956521739131, |
| "eval/E1/recall_INC": 0.3111111111111111, |
| "eval/E2/acc_on_CONSISTENT": 0.9668508287292817, |
| "eval/E2/acc_on_INCONSISTENT": 0.9710144927536232, |
| "eval/E2/accuracy": 0.968, |
| "eval/E2/f1_INC": 0.943661971830986, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9178082191780822, |
| "eval/E2/recall_INC": 0.9710144927536232, |
| "eval/E3/acc_on_CONSISTENT": 0.9464285714285714, |
| "eval/E3/acc_on_INCONSISTENT": 0.9024390243902439, |
| "eval/E3/accuracy": 0.932, |
| "eval/E3/f1_INC": 0.896969696969697, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.891566265060241, |
| "eval/E3/recall_INC": 0.9024390243902439, |
| "eval/E4/acc_on_CONSISTENT": 0.9354838709677419, |
| "eval/E4/acc_on_INCONSISTENT": 0.9052631578947369, |
| "eval/E4/accuracy": 0.924, |
| "eval/E4/f1_INC": 0.900523560209424, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8958333333333334, |
| "eval/E4/recall_INC": 0.9052631578947369, |
| "eval/edge_macro_accuracy": 0.916, |
| "eval/edge_macro_f1_INC": 0.788229983723115, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.8932038834951457, |
| "eval/overall/acc_on_INCONSISTENT": 0.891156462585034, |
| "eval/overall/accuracy": 0.892, |
| "eval/overall/f1_INC": 0.9065743944636678, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9225352112676056, |
| "eval/overall/recall_INC": 0.891156462585034, |
| "eval/overall_exact_match": 0.74, |
| "eval_label": "step_2700", |
| "step": 2700 |
| }, |
| { |
| "epoch": 0.6351446718419196, |
| "grad_norm": 0.06818749755620956, |
| "learning_rate": 3.7642493330099444e-05, |
| "loss": 0.014474105834960938, |
| "step": 2700 |
| }, |
| { |
| "epoch": 0.6398494471888967, |
| "grad_norm": 0.060922201722860336, |
| "learning_rate": 3.7157409653165173e-05, |
| "loss": 0.01806573271751404, |
| "step": 2720 |
| }, |
| { |
| "epoch": 0.6445542225358739, |
| "grad_norm": 0.07654400169849396, |
| "learning_rate": 3.66723259762309e-05, |
| "loss": 0.020049232244491576, |
| "step": 2740 |
| }, |
| { |
| "epoch": 0.6492589978828511, |
| "grad_norm": 0.047945328056812286, |
| "learning_rate": 3.618724229929663e-05, |
| "loss": 0.019580446183681488, |
| "step": 2760 |
| }, |
| { |
| "epoch": 0.6539637732298282, |
| "grad_norm": 0.16314013302326202, |
| "learning_rate": 3.570215862236236e-05, |
| "loss": 0.01818717420101166, |
| "step": 2780 |
| }, |
| { |
| "epoch": 0.6586685485768055, |
| "eval/E1/acc_on_CONSISTENT": 0.975609756097561, |
| "eval/E1/acc_on_INCONSISTENT": 0.24444444444444444, |
| "eval/E1/accuracy": 0.844, |
| "eval/E1/f1_INC": 0.360655737704918, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6875, |
| "eval/E1/recall_INC": 0.24444444444444444, |
| "eval/E2/acc_on_CONSISTENT": 0.988950276243094, |
| "eval/E2/acc_on_INCONSISTENT": 0.9710144927536232, |
| "eval/E2/accuracy": 0.984, |
| "eval/E2/f1_INC": 0.9710144927536232, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9710144927536232, |
| "eval/E2/recall_INC": 0.9710144927536232, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.9024390243902439, |
| "eval/E3/accuracy": 0.952, |
| "eval/E3/f1_INC": 0.9249999999999999, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9487179487179487, |
| "eval/E3/recall_INC": 0.9024390243902439, |
| "eval/E4/acc_on_CONSISTENT": 0.9354838709677419, |
| "eval/E4/acc_on_INCONSISTENT": 0.9157894736842105, |
| "eval/E4/accuracy": 0.928, |
| "eval/E4/f1_INC": 0.9062499999999999, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8969072164948454, |
| "eval/E4/recall_INC": 0.9157894736842105, |
| "eval/edge_macro_accuracy": 0.927, |
| "eval/edge_macro_f1_INC": 0.7907300576146352, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9223300970873787, |
| "eval/overall/acc_on_INCONSISTENT": 0.9115646258503401, |
| "eval/overall/accuracy": 0.916, |
| "eval/overall/f1_INC": 0.9273356401384083, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9436619718309859, |
| "eval/overall/recall_INC": 0.9115646258503401, |
| "eval/overall_exact_match": 0.756, |
| "eval_label": "step_2800", |
| "step": 2800 |
| }, |
| { |
| "epoch": 0.6586685485768055, |
| "grad_norm": 0.05913101136684418, |
| "learning_rate": 3.521707494542808e-05, |
| "loss": 0.019611996412277222, |
| "step": 2800 |
| }, |
| { |
| "epoch": 0.6633733239237827, |
| "grad_norm": 0.09204048663377762, |
| "learning_rate": 3.473199126849382e-05, |
| "loss": 0.021073198318481444, |
| "step": 2820 |
| }, |
| { |
| "epoch": 0.6680780992707598, |
| "grad_norm": 0.05833768472075462, |
| "learning_rate": 3.424690759155955e-05, |
| "loss": 0.01887379139661789, |
| "step": 2840 |
| }, |
| { |
| "epoch": 0.672782874617737, |
| "grad_norm": 0.10877586901187897, |
| "learning_rate": 3.376182391462527e-05, |
| "loss": 0.018790805339813234, |
| "step": 2860 |
| }, |
| { |
| "epoch": 0.6774876499647142, |
| "grad_norm": 0.040568314492702484, |
| "learning_rate": 3.3276740237691e-05, |
| "loss": 0.019745926558971404, |
| "step": 2880 |
| }, |
| { |
| "epoch": 0.6821924253116913, |
| "eval/E1/acc_on_CONSISTENT": 0.9804878048780488, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.844, |
| "eval/E1/f1_INC": 0.3389830508474576, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.7142857142857143, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.988950276243094, |
| "eval/E2/acc_on_INCONSISTENT": 0.9710144927536232, |
| "eval/E2/accuracy": 0.984, |
| "eval/E2/f1_INC": 0.9710144927536232, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9710144927536232, |
| "eval/E2/recall_INC": 0.9710144927536232, |
| "eval/E3/acc_on_CONSISTENT": 0.9821428571428571, |
| "eval/E3/acc_on_INCONSISTENT": 0.926829268292683, |
| "eval/E3/accuracy": 0.964, |
| "eval/E3/f1_INC": 0.9440993788819876, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9620253164556962, |
| "eval/E3/recall_INC": 0.926829268292683, |
| "eval/E4/acc_on_CONSISTENT": 0.967741935483871, |
| "eval/E4/acc_on_INCONSISTENT": 0.8947368421052632, |
| "eval/E4/accuracy": 0.94, |
| "eval/E4/f1_INC": 0.918918918918919, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9444444444444444, |
| "eval/E4/recall_INC": 0.8947368421052632, |
| "eval/edge_macro_accuracy": 0.9329999999999999, |
| "eval/edge_macro_f1_INC": 0.7932539603504969, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9611650485436893, |
| "eval/overall/acc_on_INCONSISTENT": 0.891156462585034, |
| "eval/overall/accuracy": 0.92, |
| "eval/overall/f1_INC": 0.9290780141843973, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9703703703703703, |
| "eval/overall/recall_INC": 0.891156462585034, |
| "eval/overall_exact_match": 0.76, |
| "eval_label": "step_2900", |
| "step": 2900 |
| }, |
| { |
| "epoch": 0.6821924253116913, |
| "grad_norm": 0.06908174604177475, |
| "learning_rate": 3.2791656560756736e-05, |
| "loss": 0.017290078103542328, |
| "step": 2900 |
| }, |
| { |
| "epoch": 0.6868972006586685, |
| "grad_norm": 0.06556062400341034, |
| "learning_rate": 3.230657288382246e-05, |
| "loss": 0.016164779663085938, |
| "step": 2920 |
| }, |
| { |
| "epoch": 0.6916019760056458, |
| "grad_norm": 0.06536762416362762, |
| "learning_rate": 3.182148920688819e-05, |
| "loss": 0.019991709291934966, |
| "step": 2940 |
| }, |
| { |
| "epoch": 0.6963067513526229, |
| "grad_norm": 0.04807380214333534, |
| "learning_rate": 3.1336405529953916e-05, |
| "loss": 0.019856654107570648, |
| "step": 2960 |
| }, |
| { |
| "epoch": 0.7010115266996001, |
| "grad_norm": 0.05577067658305168, |
| "learning_rate": 3.085132185301965e-05, |
| "loss": 0.017630717158317565, |
| "step": 2980 |
| }, |
| { |
| "epoch": 0.7057163020465773, |
| "eval/E1/acc_on_CONSISTENT": 0.9658536585365853, |
| "eval/E1/acc_on_INCONSISTENT": 0.26666666666666666, |
| "eval/E1/accuracy": 0.84, |
| "eval/E1/f1_INC": 0.37500000000000006, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.631578947368421, |
| "eval/E1/recall_INC": 0.26666666666666666, |
| "eval/E2/acc_on_CONSISTENT": 0.994475138121547, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.984, |
| "eval/E2/f1_INC": 0.9705882352941176, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9850746268656716, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9821428571428571, |
| "eval/E3/acc_on_INCONSISTENT": 0.8902439024390244, |
| "eval/E3/accuracy": 0.952, |
| "eval/E3/f1_INC": 0.9240506329113923, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9605263157894737, |
| "eval/E3/recall_INC": 0.8902439024390244, |
| "eval/E4/acc_on_CONSISTENT": 0.9548387096774194, |
| "eval/E4/acc_on_INCONSISTENT": 0.8947368421052632, |
| "eval/E4/accuracy": 0.932, |
| "eval/E4/f1_INC": 0.9090909090909092, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9239130434782609, |
| "eval/E4/recall_INC": 0.8947368421052632, |
| "eval/edge_macro_accuracy": 0.927, |
| "eval/edge_macro_f1_INC": 0.7946824443241048, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9223300970873787, |
| "eval/overall/acc_on_INCONSISTENT": 0.8843537414965986, |
| "eval/overall/accuracy": 0.9, |
| "eval/overall/f1_INC": 0.9122807017543859, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9420289855072463, |
| "eval/overall/recall_INC": 0.8843537414965986, |
| "eval/overall_exact_match": 0.748, |
| "eval_label": "step_3000", |
| "step": 3000 |
| }, |
| { |
| "epoch": 0.7057163020465773, |
| "grad_norm": 0.07990335673093796, |
| "learning_rate": 3.0366238176085378e-05, |
| "loss": 0.018373623490333557, |
| "step": 3000 |
| }, |
| { |
| "epoch": 0.7104210773935544, |
| "grad_norm": 0.042158085852861404, |
| "learning_rate": 2.9881154499151104e-05, |
| "loss": 0.01929587423801422, |
| "step": 3020 |
| }, |
| { |
| "epoch": 0.7151258527405316, |
| "grad_norm": 0.03887580707669258, |
| "learning_rate": 2.9396070822216833e-05, |
| "loss": 0.01804201304912567, |
| "step": 3040 |
| }, |
| { |
| "epoch": 0.7198306280875089, |
| "grad_norm": 0.0743173360824585, |
| "learning_rate": 2.8910987145282565e-05, |
| "loss": 0.017234998941421508, |
| "step": 3060 |
| }, |
| { |
| "epoch": 0.724535403434486, |
| "grad_norm": 0.032771673053503036, |
| "learning_rate": 2.842590346834829e-05, |
| "loss": 0.01684759259223938, |
| "step": 3080 |
| }, |
| { |
| "epoch": 0.7292401787814632, |
| "eval/E1/acc_on_CONSISTENT": 0.9317073170731708, |
| "eval/E1/acc_on_INCONSISTENT": 0.4888888888888889, |
| "eval/E1/accuracy": 0.852, |
| "eval/E1/f1_INC": 0.5432098765432098, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6111111111111112, |
| "eval/E1/recall_INC": 0.4888888888888889, |
| "eval/E2/acc_on_CONSISTENT": 0.9668508287292817, |
| "eval/E2/acc_on_INCONSISTENT": 0.9710144927536232, |
| "eval/E2/accuracy": 0.968, |
| "eval/E2/f1_INC": 0.943661971830986, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9178082191780822, |
| "eval/E2/recall_INC": 0.9710144927536232, |
| "eval/E3/acc_on_CONSISTENT": 0.9523809523809523, |
| "eval/E3/acc_on_INCONSISTENT": 0.926829268292683, |
| "eval/E3/accuracy": 0.944, |
| "eval/E3/f1_INC": 0.9156626506024096, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9047619047619048, |
| "eval/E3/recall_INC": 0.926829268292683, |
| "eval/E4/acc_on_CONSISTENT": 0.9354838709677419, |
| "eval/E4/acc_on_INCONSISTENT": 0.9052631578947369, |
| "eval/E4/accuracy": 0.924, |
| "eval/E4/f1_INC": 0.900523560209424, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8958333333333334, |
| "eval/E4/recall_INC": 0.9052631578947369, |
| "eval/edge_macro_accuracy": 0.9219999999999999, |
| "eval/edge_macro_f1_INC": 0.8257645147965074, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.912621359223301, |
| "eval/overall/acc_on_INCONSISTENT": 0.9115646258503401, |
| "eval/overall/accuracy": 0.912, |
| "eval/overall/f1_INC": 0.9241379310344827, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9370629370629371, |
| "eval/overall/recall_INC": 0.9115646258503401, |
| "eval/overall_exact_match": 0.74, |
| "eval_label": "step_3100", |
| "step": 3100 |
| }, |
| { |
| "epoch": 0.7292401787814632, |
| "grad_norm": 0.0690494254231453, |
| "learning_rate": 2.794081979141402e-05, |
| "loss": 0.020402297377586365, |
| "step": 3100 |
| }, |
| { |
| "epoch": 0.7339449541284404, |
| "grad_norm": 0.08040362596511841, |
| "learning_rate": 2.7455736114479746e-05, |
| "loss": 0.018300676345825197, |
| "step": 3120 |
| }, |
| { |
| "epoch": 0.7386497294754175, |
| "grad_norm": 0.08927752822637558, |
| "learning_rate": 2.697065243754548e-05, |
| "loss": 0.019119511544704437, |
| "step": 3140 |
| }, |
| { |
| "epoch": 0.7433545048223947, |
| "grad_norm": 0.04071735218167305, |
| "learning_rate": 2.6485568760611207e-05, |
| "loss": 0.015473829209804535, |
| "step": 3160 |
| }, |
| { |
| "epoch": 0.748059280169372, |
| "grad_norm": 0.1516721546649933, |
| "learning_rate": 2.6000485083676933e-05, |
| "loss": 0.02075496017932892, |
| "step": 3180 |
| }, |
| { |
| "epoch": 0.7527640555163491, |
| "eval/E1/acc_on_CONSISTENT": 0.9658536585365853, |
| "eval/E1/acc_on_INCONSISTENT": 0.3333333333333333, |
| "eval/E1/accuracy": 0.852, |
| "eval/E1/f1_INC": 0.4477611940298507, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6818181818181818, |
| "eval/E1/recall_INC": 0.3333333333333333, |
| "eval/E2/acc_on_CONSISTENT": 0.994475138121547, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.98, |
| "eval/E2/f1_INC": 0.962962962962963, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9848484848484849, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9702380952380952, |
| "eval/E3/acc_on_INCONSISTENT": 0.8902439024390244, |
| "eval/E3/accuracy": 0.944, |
| "eval/E3/f1_INC": 0.9125, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9358974358974359, |
| "eval/E3/recall_INC": 0.8902439024390244, |
| "eval/E4/acc_on_CONSISTENT": 0.9290322580645162, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.912, |
| "eval/E4/f1_INC": 0.8842105263157894, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8842105263157894, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.9219999999999999, |
| "eval/edge_macro_f1_INC": 0.8018586708271508, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.912621359223301, |
| "eval/overall/acc_on_INCONSISTENT": 0.8775510204081632, |
| "eval/overall/accuracy": 0.892, |
| "eval/overall/f1_INC": 0.9052631578947369, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9347826086956522, |
| "eval/overall/recall_INC": 0.8775510204081632, |
| "eval/overall_exact_match": 0.736, |
| "eval_label": "step_3200", |
| "step": 3200 |
| }, |
| { |
| "epoch": 0.7527640555163491, |
| "grad_norm": 0.0680714026093483, |
| "learning_rate": 2.5515401406742662e-05, |
| "loss": 0.018134181201457978, |
| "step": 3200 |
| }, |
| { |
| "epoch": 0.7574688308633263, |
| "grad_norm": 0.08623210340738297, |
| "learning_rate": 2.5030317729808395e-05, |
| "loss": 0.019787995517253874, |
| "step": 3220 |
| }, |
| { |
| "epoch": 0.7621736062103035, |
| "grad_norm": 0.054550524801015854, |
| "learning_rate": 2.454523405287412e-05, |
| "loss": 0.017282390594482423, |
| "step": 3240 |
| }, |
| { |
| "epoch": 0.7668783815572806, |
| "grad_norm": 0.037298593670129776, |
| "learning_rate": 2.406015037593985e-05, |
| "loss": 0.020054373145103454, |
| "step": 3260 |
| }, |
| { |
| "epoch": 0.7715831569042578, |
| "grad_norm": 0.05473620817065239, |
| "learning_rate": 2.357506669900558e-05, |
| "loss": 0.015738490223884582, |
| "step": 3280 |
| }, |
| { |
| "epoch": 0.776287932251235, |
| "eval/E1/acc_on_CONSISTENT": 0.9317073170731708, |
| "eval/E1/acc_on_INCONSISTENT": 0.4, |
| "eval/E1/accuracy": 0.836, |
| "eval/E1/f1_INC": 0.4675324675324675, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.5625, |
| "eval/E1/recall_INC": 0.4, |
| "eval/E2/acc_on_CONSISTENT": 0.9834254143646409, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.972, |
| "eval/E2/f1_INC": 0.9489051094890512, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9558823529411765, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9821428571428571, |
| "eval/E3/acc_on_INCONSISTENT": 0.9146341463414634, |
| "eval/E3/accuracy": 0.96, |
| "eval/E3/f1_INC": 0.9375000000000001, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9615384615384616, |
| "eval/E3/recall_INC": 0.9146341463414634, |
| "eval/E4/acc_on_CONSISTENT": 0.9612903225806452, |
| "eval/E4/acc_on_INCONSISTENT": 0.9157894736842105, |
| "eval/E4/accuracy": 0.944, |
| "eval/E4/f1_INC": 0.925531914893617, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9354838709677419, |
| "eval/E4/recall_INC": 0.9157894736842105, |
| "eval/edge_macro_accuracy": 0.9279999999999999, |
| "eval/edge_macro_f1_INC": 0.819867372978784, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.941747572815534, |
| "eval/overall/acc_on_INCONSISTENT": 0.891156462585034, |
| "eval/overall/accuracy": 0.912, |
| "eval/overall/f1_INC": 0.9225352112676056, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9562043795620438, |
| "eval/overall/recall_INC": 0.891156462585034, |
| "eval/overall_exact_match": 0.744, |
| "eval_label": "step_3300", |
| "step": 3300 |
| }, |
| { |
| "epoch": 0.776287932251235, |
| "grad_norm": 0.10845253616571426, |
| "learning_rate": 2.3089983022071308e-05, |
| "loss": 0.01726708859205246, |
| "step": 3300 |
| }, |
| { |
| "epoch": 0.7809927075982122, |
| "grad_norm": 0.03422887623310089, |
| "learning_rate": 2.2604899345137037e-05, |
| "loss": 0.01883469521999359, |
| "step": 3320 |
| }, |
| { |
| "epoch": 0.7856974829451894, |
| "grad_norm": 0.056281156837940216, |
| "learning_rate": 2.2119815668202766e-05, |
| "loss": 0.02208235561847687, |
| "step": 3340 |
| }, |
| { |
| "epoch": 0.7904022582921666, |
| "grad_norm": 0.057276517152786255, |
| "learning_rate": 2.1634731991268495e-05, |
| "loss": 0.016678835451602935, |
| "step": 3360 |
| }, |
| { |
| "epoch": 0.7951070336391437, |
| "grad_norm": 0.04921475797891617, |
| "learning_rate": 2.1149648314334224e-05, |
| "loss": 0.016811944544315338, |
| "step": 3380 |
| }, |
| { |
| "epoch": 0.7998118089861209, |
| "eval/E1/acc_on_CONSISTENT": 0.9707317073170731, |
| "eval/E1/acc_on_INCONSISTENT": 0.28888888888888886, |
| "eval/E1/accuracy": 0.848, |
| "eval/E1/f1_INC": 0.40625, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6842105263157895, |
| "eval/E1/recall_INC": 0.28888888888888886, |
| "eval/E2/acc_on_CONSISTENT": 0.988950276243094, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.976, |
| "eval/E2/f1_INC": 0.9558823529411764, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9701492537313433, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9642857142857143, |
| "eval/E3/acc_on_INCONSISTENT": 0.9024390243902439, |
| "eval/E3/accuracy": 0.944, |
| "eval/E3/f1_INC": 0.9135802469135802, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.925, |
| "eval/E3/recall_INC": 0.9024390243902439, |
| "eval/E4/acc_on_CONSISTENT": 0.967741935483871, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.936, |
| "eval/E4/f1_INC": 0.9130434782608695, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9438202247191011, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.9259999999999999, |
| "eval/edge_macro_f1_INC": 0.7971890195289065, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9223300970873787, |
| "eval/overall/acc_on_INCONSISTENT": 0.8843537414965986, |
| "eval/overall/accuracy": 0.9, |
| "eval/overall/f1_INC": 0.9122807017543859, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9420289855072463, |
| "eval/overall/recall_INC": 0.8843537414965986, |
| "eval/overall_exact_match": 0.74, |
| "eval_label": "step_3400", |
| "step": 3400 |
| }, |
| { |
| "epoch": 0.7998118089861209, |
| "grad_norm": 0.11905638128519058, |
| "learning_rate": 2.0664564637399954e-05, |
| "loss": 0.015251293778419495, |
| "step": 3400 |
| }, |
| { |
| "epoch": 0.8045165843330981, |
| "grad_norm": 0.04575642943382263, |
| "learning_rate": 2.0179480960465683e-05, |
| "loss": 0.020809194445610045, |
| "step": 3420 |
| }, |
| { |
| "epoch": 0.8092213596800752, |
| "grad_norm": 0.044995591044425964, |
| "learning_rate": 1.969439728353141e-05, |
| "loss": 0.014254528284072875, |
| "step": 3440 |
| }, |
| { |
| "epoch": 0.8139261350270525, |
| "grad_norm": 0.05187941715121269, |
| "learning_rate": 1.920931360659714e-05, |
| "loss": 0.01632629781961441, |
| "step": 3460 |
| }, |
| { |
| "epoch": 0.8186309103740297, |
| "grad_norm": 0.02889666147530079, |
| "learning_rate": 1.8724229929662867e-05, |
| "loss": 0.020733726024627686, |
| "step": 3480 |
| }, |
| { |
| "epoch": 0.8233356857210068, |
| "eval/E1/acc_on_CONSISTENT": 0.9463414634146341, |
| "eval/E1/acc_on_INCONSISTENT": 0.4222222222222222, |
| "eval/E1/accuracy": 0.852, |
| "eval/E1/f1_INC": 0.5066666666666666, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6333333333333333, |
| "eval/E1/recall_INC": 0.4222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.9723756906077348, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.968, |
| "eval/E2/f1_INC": 0.9428571428571428, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9295774647887324, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9523809523809523, |
| "eval/E3/acc_on_INCONSISTENT": 0.9390243902439024, |
| "eval/E3/accuracy": 0.948, |
| "eval/E3/f1_INC": 0.9221556886227545, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9058823529411765, |
| "eval/E3/recall_INC": 0.9390243902439024, |
| "eval/E4/acc_on_CONSISTENT": 0.9354838709677419, |
| "eval/E4/acc_on_INCONSISTENT": 0.9052631578947369, |
| "eval/E4/accuracy": 0.924, |
| "eval/E4/f1_INC": 0.900523560209424, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.8958333333333334, |
| "eval/E4/recall_INC": 0.9052631578947369, |
| "eval/edge_macro_accuracy": 0.923, |
| "eval/edge_macro_f1_INC": 0.818050764588997, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9029126213592233, |
| "eval/overall/acc_on_INCONSISTENT": 0.9047619047619048, |
| "eval/overall/accuracy": 0.904, |
| "eval/overall/f1_INC": 0.9172413793103449, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9300699300699301, |
| "eval/overall/recall_INC": 0.9047619047619048, |
| "eval/overall_exact_match": 0.74, |
| "eval_label": "step_3500", |
| "step": 3500 |
| }, |
| { |
| "epoch": 0.8233356857210068, |
| "grad_norm": 0.05501768738031387, |
| "learning_rate": 1.82391462527286e-05, |
| "loss": 0.01859801113605499, |
| "step": 3500 |
| }, |
| { |
| "epoch": 0.828040461067984, |
| "grad_norm": 0.06756485998630524, |
| "learning_rate": 1.7754062575794325e-05, |
| "loss": 0.017083537578582764, |
| "step": 3520 |
| }, |
| { |
| "epoch": 0.8327452364149612, |
| "grad_norm": 0.05211762711405754, |
| "learning_rate": 1.7268978898860054e-05, |
| "loss": 0.016947761178016663, |
| "step": 3540 |
| }, |
| { |
| "epoch": 0.8374500117619383, |
| "grad_norm": 0.053344082087278366, |
| "learning_rate": 1.6783895221925783e-05, |
| "loss": 0.02057153731584549, |
| "step": 3560 |
| }, |
| { |
| "epoch": 0.8421547871089156, |
| "grad_norm": 0.12314067780971527, |
| "learning_rate": 1.6298811544991512e-05, |
| "loss": 0.017026612162590028, |
| "step": 3580 |
| }, |
| { |
| "epoch": 0.8468595624558928, |
| "eval/E1/acc_on_CONSISTENT": 0.9609756097560975, |
| "eval/E1/acc_on_INCONSISTENT": 0.35555555555555557, |
| "eval/E1/accuracy": 0.852, |
| "eval/E1/f1_INC": 0.46376811594202905, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6666666666666666, |
| "eval/E1/recall_INC": 0.35555555555555557, |
| "eval/E2/acc_on_CONSISTENT": 0.9834254143646409, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.972, |
| "eval/E2/f1_INC": 0.9489051094890512, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9558823529411765, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9702380952380952, |
| "eval/E3/acc_on_INCONSISTENT": 0.8780487804878049, |
| "eval/E3/accuracy": 0.94, |
| "eval/E3/f1_INC": 0.9056603773584906, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.935064935064935, |
| "eval/E3/recall_INC": 0.8780487804878049, |
| "eval/E4/acc_on_CONSISTENT": 0.9483870967741935, |
| "eval/E4/acc_on_INCONSISTENT": 0.8842105263157894, |
| "eval/E4/accuracy": 0.924, |
| "eval/E4/f1_INC": 0.8983957219251336, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9130434782608695, |
| "eval/E4/recall_INC": 0.8842105263157894, |
| "eval/edge_macro_accuracy": 0.9219999999999999, |
| "eval/edge_macro_f1_INC": 0.8041823311786761, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9223300970873787, |
| "eval/overall/acc_on_INCONSISTENT": 0.8707482993197279, |
| "eval/overall/accuracy": 0.892, |
| "eval/overall/f1_INC": 0.9045936395759717, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9411764705882353, |
| "eval/overall/recall_INC": 0.8707482993197279, |
| "eval/overall_exact_match": 0.732, |
| "eval_label": "step_3600", |
| "step": 3600 |
| }, |
| { |
| "epoch": 0.8468595624558928, |
| "grad_norm": 0.058041803538799286, |
| "learning_rate": 1.581372786805724e-05, |
| "loss": 0.01929272264242172, |
| "step": 3600 |
| }, |
| { |
| "epoch": 0.8515643378028699, |
| "grad_norm": 0.05177493393421173, |
| "learning_rate": 1.532864419112297e-05, |
| "loss": 0.01702711880207062, |
| "step": 3620 |
| }, |
| { |
| "epoch": 0.8562691131498471, |
| "grad_norm": 0.07954961806535721, |
| "learning_rate": 1.4843560514188698e-05, |
| "loss": 0.018577194213867186, |
| "step": 3640 |
| }, |
| { |
| "epoch": 0.8609738884968243, |
| "grad_norm": 0.06260311603546143, |
| "learning_rate": 1.4358476837254429e-05, |
| "loss": 0.01722830384969711, |
| "step": 3660 |
| }, |
| { |
| "epoch": 0.8656786638438014, |
| "grad_norm": 0.0682927593588829, |
| "learning_rate": 1.3873393160320156e-05, |
| "loss": 0.017568738758563997, |
| "step": 3680 |
| }, |
| { |
| "epoch": 0.8703834391907787, |
| "eval/E1/acc_on_CONSISTENT": 0.9902439024390244, |
| "eval/E1/acc_on_INCONSISTENT": 0.26666666666666666, |
| "eval/E1/accuracy": 0.86, |
| "eval/E1/f1_INC": 0.4067796610169491, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.8571428571428571, |
| "eval/E1/recall_INC": 0.26666666666666666, |
| "eval/E2/acc_on_CONSISTENT": 0.988950276243094, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.976, |
| "eval/E2/f1_INC": 0.9558823529411764, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9701492537313433, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.8780487804878049, |
| "eval/E3/accuracy": 0.944, |
| "eval/E3/f1_INC": 0.9113924050632912, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9473684210526315, |
| "eval/E3/recall_INC": 0.8780487804878049, |
| "eval/E4/acc_on_CONSISTENT": 0.967741935483871, |
| "eval/E4/acc_on_INCONSISTENT": 0.8947368421052632, |
| "eval/E4/accuracy": 0.94, |
| "eval/E4/f1_INC": 0.918918918918919, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9444444444444444, |
| "eval/E4/recall_INC": 0.8947368421052632, |
| "eval/edge_macro_accuracy": 0.9299999999999999, |
| "eval/edge_macro_f1_INC": 0.7982433344850839, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9611650485436893, |
| "eval/overall/acc_on_INCONSISTENT": 0.8707482993197279, |
| "eval/overall/accuracy": 0.908, |
| "eval/overall/f1_INC": 0.9175627240143368, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9696969696969697, |
| "eval/overall/recall_INC": 0.8707482993197279, |
| "eval/overall_exact_match": 0.768, |
| "eval_label": "step_3700", |
| "step": 3700 |
| }, |
| { |
| "epoch": 0.8703834391907787, |
| "grad_norm": 0.04631944000720978, |
| "learning_rate": 1.3388309483385885e-05, |
| "loss": 0.01513085961341858, |
| "step": 3700 |
| }, |
| { |
| "epoch": 0.8750882145377559, |
| "grad_norm": 0.09454777091741562, |
| "learning_rate": 1.2903225806451613e-05, |
| "loss": 0.019655285775661467, |
| "step": 3720 |
| }, |
| { |
| "epoch": 0.879792989884733, |
| "grad_norm": 0.08288902789354324, |
| "learning_rate": 1.2418142129517342e-05, |
| "loss": 0.016171494126319887, |
| "step": 3740 |
| }, |
| { |
| "epoch": 0.8844977652317102, |
| "grad_norm": 0.14061884582042694, |
| "learning_rate": 1.1933058452583071e-05, |
| "loss": 0.018433985114097596, |
| "step": 3760 |
| }, |
| { |
| "epoch": 0.8892025405786874, |
| "grad_norm": 0.06681777536869049, |
| "learning_rate": 1.14479747756488e-05, |
| "loss": 0.02021569013595581, |
| "step": 3780 |
| }, |
| { |
| "epoch": 0.8939073159256645, |
| "eval/E1/acc_on_CONSISTENT": 0.9902439024390244, |
| "eval/E1/acc_on_INCONSISTENT": 0.2222222222222222, |
| "eval/E1/accuracy": 0.852, |
| "eval/E1/f1_INC": 0.3508771929824561, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.8333333333333334, |
| "eval/E1/recall_INC": 0.2222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.988950276243094, |
| "eval/E2/acc_on_INCONSISTENT": 0.927536231884058, |
| "eval/E2/accuracy": 0.972, |
| "eval/E2/f1_INC": 0.9481481481481481, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9696969696969697, |
| "eval/E2/recall_INC": 0.927536231884058, |
| "eval/E3/acc_on_CONSISTENT": 0.9821428571428571, |
| "eval/E3/acc_on_INCONSISTENT": 0.8658536585365854, |
| "eval/E3/accuracy": 0.944, |
| "eval/E3/f1_INC": 0.9102564102564101, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9594594594594594, |
| "eval/E3/recall_INC": 0.8658536585365854, |
| "eval/E4/acc_on_CONSISTENT": 0.9741935483870968, |
| "eval/E4/acc_on_INCONSISTENT": 0.8947368421052632, |
| "eval/E4/accuracy": 0.944, |
| "eval/E4/f1_INC": 0.9239130434782609, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9550561797752809, |
| "eval/E4/recall_INC": 0.8947368421052632, |
| "eval/edge_macro_accuracy": 0.9279999999999999, |
| "eval/edge_macro_f1_INC": 0.7832986987163189, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.970873786407767, |
| "eval/overall/acc_on_INCONSISTENT": 0.8639455782312925, |
| "eval/overall/accuracy": 0.908, |
| "eval/overall/f1_INC": 0.9169675090252707, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9769230769230769, |
| "eval/overall/recall_INC": 0.8639455782312925, |
| "eval/overall_exact_match": 0.768, |
| "eval_label": "step_3800", |
| "step": 3800 |
| }, |
| { |
| "epoch": 0.8939073159256645, |
| "grad_norm": 0.06884897500276566, |
| "learning_rate": 1.0962891098714528e-05, |
| "loss": 0.017668071389198303, |
| "step": 3800 |
| }, |
| { |
| "epoch": 0.8986120912726417, |
| "grad_norm": 0.08243655413389206, |
| "learning_rate": 1.0477807421780257e-05, |
| "loss": 0.02169010043144226, |
| "step": 3820 |
| }, |
| { |
| "epoch": 0.903316866619619, |
| "grad_norm": 0.06311216950416565, |
| "learning_rate": 9.992723744845986e-06, |
| "loss": 0.018402168154716493, |
| "step": 3840 |
| }, |
| { |
| "epoch": 0.9080216419665961, |
| "grad_norm": 0.046876635402441025, |
| "learning_rate": 9.507640067911715e-06, |
| "loss": 0.012961253523826599, |
| "step": 3860 |
| }, |
| { |
| "epoch": 0.9127264173135733, |
| "grad_norm": 0.058287572115659714, |
| "learning_rate": 9.022556390977444e-06, |
| "loss": 0.01640145927667618, |
| "step": 3880 |
| }, |
| { |
| "epoch": 0.9174311926605505, |
| "eval/E1/acc_on_CONSISTENT": 0.9707317073170731, |
| "eval/E1/acc_on_INCONSISTENT": 0.4222222222222222, |
| "eval/E1/accuracy": 0.872, |
| "eval/E1/f1_INC": 0.5428571428571428, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.76, |
| "eval/E1/recall_INC": 0.4222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.994475138121547, |
| "eval/E2/acc_on_INCONSISTENT": 0.9420289855072463, |
| "eval/E2/accuracy": 0.98, |
| "eval/E2/f1_INC": 0.962962962962963, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9848484848484849, |
| "eval/E2/recall_INC": 0.9420289855072463, |
| "eval/E3/acc_on_CONSISTENT": 0.9523809523809523, |
| "eval/E3/acc_on_INCONSISTENT": 0.8902439024390244, |
| "eval/E3/accuracy": 0.932, |
| "eval/E3/f1_INC": 0.8957055214723927, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9012345679012346, |
| "eval/E3/recall_INC": 0.8902439024390244, |
| "eval/E4/acc_on_CONSISTENT": 0.9612903225806452, |
| "eval/E4/acc_on_INCONSISTENT": 0.8947368421052632, |
| "eval/E4/accuracy": 0.936, |
| "eval/E4/f1_INC": 0.913978494623656, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9340659340659341, |
| "eval/E4/recall_INC": 0.8947368421052632, |
| "eval/edge_macro_accuracy": 0.93, |
| "eval/edge_macro_f1_INC": 0.8288760304790386, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.9320388349514563, |
| "eval/overall/acc_on_INCONSISTENT": 0.8707482993197279, |
| "eval/overall/accuracy": 0.896, |
| "eval/overall/f1_INC": 0.9078014184397163, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9481481481481482, |
| "eval/overall/recall_INC": 0.8707482993197279, |
| "eval/overall_exact_match": 0.764, |
| "eval_label": "step_3900", |
| "step": 3900 |
| }, |
| { |
| "epoch": 0.9174311926605505, |
| "grad_norm": 0.04448636621236801, |
| "learning_rate": 8.537472714043173e-06, |
| "loss": 0.01752175986766815, |
| "step": 3900 |
| }, |
| { |
| "epoch": 0.9221359680075276, |
| "grad_norm": 0.10632307827472687, |
| "learning_rate": 8.0523890371089e-06, |
| "loss": 0.020644263923168184, |
| "step": 3920 |
| }, |
| { |
| "epoch": 0.9268407433545048, |
| "grad_norm": 0.04823046177625656, |
| "learning_rate": 7.56730536017463e-06, |
| "loss": 0.017941384017467497, |
| "step": 3940 |
| }, |
| { |
| "epoch": 0.9315455187014821, |
| "grad_norm": 0.029647739604115486, |
| "learning_rate": 7.082221683240359e-06, |
| "loss": 0.017719915509223937, |
| "step": 3960 |
| }, |
| { |
| "epoch": 0.9362502940484592, |
| "grad_norm": 0.0772344246506691, |
| "learning_rate": 6.597138006306087e-06, |
| "loss": 0.018131235241889955, |
| "step": 3980 |
| }, |
| { |
| "epoch": 0.9409550693954364, |
| "eval/E1/acc_on_CONSISTENT": 0.9707317073170731, |
| "eval/E1/acc_on_INCONSISTENT": 0.4222222222222222, |
| "eval/E1/accuracy": 0.872, |
| "eval/E1/f1_INC": 0.5428571428571428, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.76, |
| "eval/E1/recall_INC": 0.4222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.9779005524861878, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.972, |
| "eval/E2/f1_INC": 0.9496402877697843, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9428571428571428, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9761904761904762, |
| "eval/E3/acc_on_INCONSISTENT": 0.926829268292683, |
| "eval/E3/accuracy": 0.96, |
| "eval/E3/f1_INC": 0.9382716049382716, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.95, |
| "eval/E3/recall_INC": 0.926829268292683, |
| "eval/E4/acc_on_CONSISTENT": 0.9612903225806452, |
| "eval/E4/acc_on_INCONSISTENT": 0.9157894736842105, |
| "eval/E4/accuracy": 0.944, |
| "eval/E4/f1_INC": 0.925531914893617, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9354838709677419, |
| "eval/E4/recall_INC": 0.9157894736842105, |
| "eval/edge_macro_accuracy": 0.9369999999999999, |
| "eval/edge_macro_f1_INC": 0.8390752376147039, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.941747572815534, |
| "eval/overall/acc_on_INCONSISTENT": 0.8979591836734694, |
| "eval/overall/accuracy": 0.916, |
| "eval/overall/f1_INC": 0.9263157894736843, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9565217391304348, |
| "eval/overall/recall_INC": 0.8979591836734694, |
| "eval/overall_exact_match": 0.784, |
| "eval_label": "step_4000", |
| "step": 4000 |
| }, |
| { |
| "epoch": 0.9409550693954364, |
| "grad_norm": 0.05973908305168152, |
| "learning_rate": 6.112054329371817e-06, |
| "loss": 0.01839745044708252, |
| "step": 4000 |
| }, |
| { |
| "epoch": 0.9456598447424136, |
| "grad_norm": 0.060186609625816345, |
| "learning_rate": 5.626970652437546e-06, |
| "loss": 0.020127423107624054, |
| "step": 4020 |
| }, |
| { |
| "epoch": 0.9503646200893907, |
| "grad_norm": 0.08561345934867859, |
| "learning_rate": 5.1418869755032745e-06, |
| "loss": 0.01490895301103592, |
| "step": 4040 |
| }, |
| { |
| "epoch": 0.9550693954363679, |
| "grad_norm": 0.028145257383584976, |
| "learning_rate": 4.656803298569004e-06, |
| "loss": 0.017035940289497377, |
| "step": 4060 |
| }, |
| { |
| "epoch": 0.9597741707833451, |
| "grad_norm": 0.09954910725355148, |
| "learning_rate": 4.171719621634732e-06, |
| "loss": 0.01837383508682251, |
| "step": 4080 |
| }, |
| { |
| "epoch": 0.9644789461303223, |
| "eval/E1/acc_on_CONSISTENT": 0.9512195121951219, |
| "eval/E1/acc_on_INCONSISTENT": 0.4444444444444444, |
| "eval/E1/accuracy": 0.86, |
| "eval/E1/f1_INC": 0.5333333333333333, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.6666666666666666, |
| "eval/E1/recall_INC": 0.4444444444444444, |
| "eval/E2/acc_on_CONSISTENT": 0.9779005524861878, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.972, |
| "eval/E2/f1_INC": 0.9496402877697843, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9428571428571428, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9702380952380952, |
| "eval/E3/acc_on_INCONSISTENT": 0.926829268292683, |
| "eval/E3/accuracy": 0.956, |
| "eval/E3/f1_INC": 0.9325153374233128, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9382716049382716, |
| "eval/E3/recall_INC": 0.926829268292683, |
| "eval/E4/acc_on_CONSISTENT": 0.9612903225806452, |
| "eval/E4/acc_on_INCONSISTENT": 0.9157894736842105, |
| "eval/E4/accuracy": 0.944, |
| "eval/E4/f1_INC": 0.925531914893617, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.9354838709677419, |
| "eval/E4/recall_INC": 0.9157894736842105, |
| "eval/edge_macro_accuracy": 0.9329999999999999, |
| "eval/edge_macro_f1_INC": 0.8352552183550118, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.941747572815534, |
| "eval/overall/acc_on_INCONSISTENT": 0.8979591836734694, |
| "eval/overall/accuracy": 0.916, |
| "eval/overall/f1_INC": 0.9263157894736843, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9565217391304348, |
| "eval/overall/recall_INC": 0.8979591836734694, |
| "eval/overall_exact_match": 0.772, |
| "eval_label": "step_4100", |
| "step": 4100 |
| }, |
| { |
| "epoch": 0.9644789461303223, |
| "grad_norm": 0.06335736811161041, |
| "learning_rate": 3.686635944700461e-06, |
| "loss": 0.018984177708625795, |
| "step": 4100 |
| }, |
| { |
| "epoch": 0.9691837214772995, |
| "grad_norm": 0.06948219239711761, |
| "learning_rate": 3.2015522677661897e-06, |
| "loss": 0.01891358643770218, |
| "step": 4120 |
| }, |
| { |
| "epoch": 0.9738884968242767, |
| "grad_norm": 0.04544758051633835, |
| "learning_rate": 2.7164685908319184e-06, |
| "loss": 0.016408444941043855, |
| "step": 4140 |
| }, |
| { |
| "epoch": 0.9785932721712538, |
| "grad_norm": 0.05341501906514168, |
| "learning_rate": 2.2313849138976476e-06, |
| "loss": 0.01944984793663025, |
| "step": 4160 |
| }, |
| { |
| "epoch": 0.983298047518231, |
| "grad_norm": 0.054112840443849564, |
| "learning_rate": 1.7463012369633763e-06, |
| "loss": 0.020673635601997375, |
| "step": 4180 |
| }, |
| { |
| "epoch": 0.9880028228652082, |
| "eval/E1/acc_on_CONSISTENT": 0.9707317073170731, |
| "eval/E1/acc_on_INCONSISTENT": 0.4222222222222222, |
| "eval/E1/accuracy": 0.872, |
| "eval/E1/f1_INC": 0.5428571428571428, |
| "eval/E1/n": 250.0, |
| "eval/E1/precision_INC": 0.76, |
| "eval/E1/recall_INC": 0.4222222222222222, |
| "eval/E2/acc_on_CONSISTENT": 0.988950276243094, |
| "eval/E2/acc_on_INCONSISTENT": 0.9565217391304348, |
| "eval/E2/accuracy": 0.98, |
| "eval/E2/f1_INC": 0.9635036496350365, |
| "eval/E2/n": 250.0, |
| "eval/E2/precision_INC": 0.9705882352941176, |
| "eval/E2/recall_INC": 0.9565217391304348, |
| "eval/E3/acc_on_CONSISTENT": 0.9702380952380952, |
| "eval/E3/acc_on_INCONSISTENT": 0.9390243902439024, |
| "eval/E3/accuracy": 0.96, |
| "eval/E3/f1_INC": 0.9390243902439024, |
| "eval/E3/n": 250.0, |
| "eval/E3/precision_INC": 0.9390243902439024, |
| "eval/E3/recall_INC": 0.9390243902439024, |
| "eval/E4/acc_on_CONSISTENT": 0.967741935483871, |
| "eval/E4/acc_on_INCONSISTENT": 0.9052631578947369, |
| "eval/E4/accuracy": 0.944, |
| "eval/E4/f1_INC": 0.924731182795699, |
| "eval/E4/n": 250.0, |
| "eval/E4/precision_INC": 0.945054945054945, |
| "eval/E4/recall_INC": 0.9052631578947369, |
| "eval/edge_macro_accuracy": 0.939, |
| "eval/edge_macro_f1_INC": 0.8425290913829452, |
| "eval/malformed_rate": 0.0, |
| "eval/n_eval": 1000.0, |
| "eval/overall/acc_on_CONSISTENT": 0.941747572815534, |
| "eval/overall/acc_on_INCONSISTENT": 0.891156462585034, |
| "eval/overall/accuracy": 0.912, |
| "eval/overall/f1_INC": 0.9225352112676056, |
| "eval/overall/n": 250.0, |
| "eval/overall/n_complete_groups": 250.0, |
| "eval/overall/precision_INC": 0.9562043795620438, |
| "eval/overall/recall_INC": 0.891156462585034, |
| "eval/overall_exact_match": 0.78, |
| "eval_label": "step_4200", |
| "step": 4200 |
| }, |
| { |
| "epoch": 0.9880028228652082, |
| "grad_norm": 0.04273684322834015, |
| "learning_rate": 1.2612175600291052e-06, |
| "loss": 0.018599729239940643, |
| "step": 4200 |
| } |
| ], |
| "logging_steps": 20, |
| "max_steps": 4251, |
| "num_input_tokens_seen": 0, |
| "num_train_epochs": 1, |
| "save_steps": 300, |
| "stateful_callbacks": { |
| "TrainerControl": { |
| "args": { |
| "should_epoch_stop": false, |
| "should_evaluate": false, |
| "should_log": false, |
| "should_save": true, |
| "should_training_stop": false |
| }, |
| "attributes": {} |
| } |
| }, |
| "total_flos": 4.197003788524978e+18, |
| "train_batch_size": 8, |
| "trial_name": null, |
| "trial_params": null |
| } |
|
|
