1093 lines
37 KiB
JSON
1093 lines
37 KiB
JSON
{
|
|
"best_global_step": 21,
|
|
"best_metric": 0.6508771929824562,
|
|
"best_model_checkpoint": "/root/ARC-Easy_Llama-3.2-1B/no_sweep/oqrx1b71_lunar-oath-85/checkpoint-21",
|
|
"epoch": 21.0,
|
|
"eval_steps": 500,
|
|
"global_step": 21,
|
|
"is_hyper_param_search": false,
|
|
"is_local_process_zero": true,
|
|
"is_world_process_zero": true,
|
|
"log_history": [
|
|
{
|
|
"epoch": 0,
|
|
"eval_accumulated_loss": 875.1691761016846,
|
|
"eval_accuracy": 0.3017543859649123,
|
|
"eval_accuracy_32": 0.9746835443037974,
|
|
"eval_accuracy_33": 0.0,
|
|
"eval_accuracy_34": 0.11971830985915492,
|
|
"eval_accuracy_35": 0.00847457627118644,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 170.0,
|
|
"eval_correct_gen_preds_32": 154.0,
|
|
"eval_correct_gen_preds_33": 0.0,
|
|
"eval_correct_gen_preds_34": 15.0,
|
|
"eval_correct_gen_preds_35": 1.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 172.0,
|
|
"eval_correct_preds_32": 154.0,
|
|
"eval_correct_preds_33": 0.0,
|
|
"eval_correct_preds_34": 17.0,
|
|
"eval_correct_preds_35": 1.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.2982456140350877,
|
|
"eval_gen_accuracy_32": 0.9746835443037974,
|
|
"eval_gen_accuracy_33": 0.0,
|
|
"eval_gen_accuracy_34": 0.1056338028169014,
|
|
"eval_gen_accuracy_35": 0.00847457627118644,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 1.5353845357894897,
|
|
"eval_mdl": 1262.6022303007803,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.0141,
|
|
"eval_samples_per_second": 283.004,
|
|
"eval_steps_per_second": 2.979,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 0
|
|
},
|
|
{
|
|
"epoch": 1.0,
|
|
"grad_norm": 117.0,
|
|
"learning_rate": 0.0,
|
|
"loss": 1.4592,
|
|
"step": 1
|
|
},
|
|
{
|
|
"epoch": 1.0,
|
|
"eval_accumulated_loss": 875.1691761016846,
|
|
"eval_accuracy": 0.3017543859649123,
|
|
"eval_accuracy_32": 0.9746835443037974,
|
|
"eval_accuracy_33": 0.0,
|
|
"eval_accuracy_34": 0.11971830985915492,
|
|
"eval_accuracy_35": 0.00847457627118644,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 170.0,
|
|
"eval_correct_gen_preds_32": 154.0,
|
|
"eval_correct_gen_preds_33": 0.0,
|
|
"eval_correct_gen_preds_34": 15.0,
|
|
"eval_correct_gen_preds_35": 1.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 172.0,
|
|
"eval_correct_preds_32": 154.0,
|
|
"eval_correct_preds_33": 0.0,
|
|
"eval_correct_preds_34": 17.0,
|
|
"eval_correct_preds_35": 1.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.2982456140350877,
|
|
"eval_gen_accuracy_32": 0.9746835443037974,
|
|
"eval_gen_accuracy_33": 0.0,
|
|
"eval_gen_accuracy_34": 0.1056338028169014,
|
|
"eval_gen_accuracy_35": 0.00847457627118644,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 1.5353845357894897,
|
|
"eval_mdl": 1262.6022303007803,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 1.9316,
|
|
"eval_samples_per_second": 295.093,
|
|
"eval_steps_per_second": 3.106,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 1
|
|
},
|
|
{
|
|
"epoch": 2.0,
|
|
"grad_norm": 116.0,
|
|
"learning_rate": 2e-05,
|
|
"loss": 1.4587,
|
|
"step": 2
|
|
},
|
|
{
|
|
"epoch": 2.0,
|
|
"eval_accumulated_loss": 1628.622148513794,
|
|
"eval_accuracy": 0.3298245614035088,
|
|
"eval_accuracy_32": 0.0,
|
|
"eval_accuracy_33": 0.3026315789473684,
|
|
"eval_accuracy_34": 0.9929577464788732,
|
|
"eval_accuracy_35": 0.00847457627118644,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 188.0,
|
|
"eval_correct_gen_preds_32": 0.0,
|
|
"eval_correct_gen_preds_33": 46.0,
|
|
"eval_correct_gen_preds_34": 141.0,
|
|
"eval_correct_gen_preds_35": 1.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 188.0,
|
|
"eval_correct_preds_32": 0.0,
|
|
"eval_correct_preds_33": 46.0,
|
|
"eval_correct_preds_34": 141.0,
|
|
"eval_correct_preds_35": 1.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.3298245614035088,
|
|
"eval_gen_accuracy_32": 0.0,
|
|
"eval_gen_accuracy_33": 0.3026315789473684,
|
|
"eval_gen_accuracy_34": 0.9929577464788732,
|
|
"eval_gen_accuracy_35": 0.00847457627118644,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 2.857232093811035,
|
|
"eval_mdl": 2349.6050971427794,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 1.9821,
|
|
"eval_samples_per_second": 287.57,
|
|
"eval_steps_per_second": 3.027,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 2
|
|
},
|
|
{
|
|
"epoch": 3.0,
|
|
"grad_norm": 125.0,
|
|
"learning_rate": 1.9994965423831853e-05,
|
|
"loss": 1.7615,
|
|
"step": 3
|
|
},
|
|
{
|
|
"epoch": 3.0,
|
|
"eval_accumulated_loss": 863.5332918167114,
|
|
"eval_accuracy": 0.30350877192982456,
|
|
"eval_accuracy_32": 0.0,
|
|
"eval_accuracy_33": 0.993421052631579,
|
|
"eval_accuracy_34": 0.056338028169014086,
|
|
"eval_accuracy_35": 0.11864406779661017,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 173.0,
|
|
"eval_correct_gen_preds_32": 0.0,
|
|
"eval_correct_gen_preds_33": 151.0,
|
|
"eval_correct_gen_preds_34": 8.0,
|
|
"eval_correct_gen_preds_35": 14.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 173.0,
|
|
"eval_correct_preds_32": 0.0,
|
|
"eval_correct_preds_33": 151.0,
|
|
"eval_correct_preds_34": 8.0,
|
|
"eval_correct_preds_35": 14.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.30350877192982456,
|
|
"eval_gen_accuracy_32": 0.0,
|
|
"eval_gen_accuracy_33": 0.993421052631579,
|
|
"eval_gen_accuracy_34": 0.056338028169014086,
|
|
"eval_gen_accuracy_35": 0.11864406779661017,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 1.5149707794189453,
|
|
"eval_mdl": 1245.8151977464918,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 1.8611,
|
|
"eval_samples_per_second": 306.269,
|
|
"eval_steps_per_second": 3.224,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 3
|
|
},
|
|
{
|
|
"epoch": 4.0,
|
|
"grad_norm": 85.0,
|
|
"learning_rate": 1.9979866764718846e-05,
|
|
"loss": 0.683,
|
|
"step": 4
|
|
},
|
|
{
|
|
"epoch": 4.0,
|
|
"eval_accumulated_loss": 786.9393863677979,
|
|
"eval_accuracy": 0.5385964912280702,
|
|
"eval_accuracy_32": 0.37341772151898733,
|
|
"eval_accuracy_33": 0.8289473684210527,
|
|
"eval_accuracy_34": 0.5915492957746479,
|
|
"eval_accuracy_35": 0.3220338983050847,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 254.0,
|
|
"eval_correct_gen_preds_32": 39.0,
|
|
"eval_correct_gen_preds_33": 107.0,
|
|
"eval_correct_gen_preds_34": 76.0,
|
|
"eval_correct_gen_preds_35": 32.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 307.0,
|
|
"eval_correct_preds_32": 59.0,
|
|
"eval_correct_preds_33": 126.0,
|
|
"eval_correct_preds_34": 84.0,
|
|
"eval_correct_preds_35": 38.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.4456140350877193,
|
|
"eval_gen_accuracy_32": 0.2468354430379747,
|
|
"eval_gen_accuracy_33": 0.7039473684210527,
|
|
"eval_gen_accuracy_34": 0.5352112676056338,
|
|
"eval_gen_accuracy_35": 0.2711864406779661,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 1.3805955648422241,
|
|
"eval_mdl": 1135.313550193026,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 3.3581,
|
|
"eval_samples_per_second": 169.741,
|
|
"eval_steps_per_second": 1.787,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 4
|
|
},
|
|
{
|
|
"epoch": 5.0,
|
|
"grad_norm": 14.125,
|
|
"learning_rate": 1.9954719225730847e-05,
|
|
"loss": 0.0591,
|
|
"step": 5
|
|
},
|
|
{
|
|
"epoch": 5.0,
|
|
"eval_accumulated_loss": 1043.7736988067627,
|
|
"eval_accuracy": 0.624561403508772,
|
|
"eval_accuracy_32": 0.620253164556962,
|
|
"eval_accuracy_33": 0.756578947368421,
|
|
"eval_accuracy_34": 0.647887323943662,
|
|
"eval_accuracy_35": 0.4322033898305085,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 262.0,
|
|
"eval_correct_gen_preds_32": 53.0,
|
|
"eval_correct_gen_preds_33": 98.0,
|
|
"eval_correct_gen_preds_34": 69.0,
|
|
"eval_correct_gen_preds_35": 42.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 356.0,
|
|
"eval_correct_preds_32": 98.0,
|
|
"eval_correct_preds_33": 115.0,
|
|
"eval_correct_preds_34": 92.0,
|
|
"eval_correct_preds_35": 51.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.45964912280701753,
|
|
"eval_gen_accuracy_32": 0.33544303797468356,
|
|
"eval_gen_accuracy_33": 0.6447368421052632,
|
|
"eval_gen_accuracy_34": 0.4859154929577465,
|
|
"eval_gen_accuracy_35": 0.3559322033898305,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 1.8311821222305298,
|
|
"eval_mdl": 1505.847139078847,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.0813,
|
|
"eval_samples_per_second": 273.865,
|
|
"eval_steps_per_second": 2.883,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 5
|
|
},
|
|
{
|
|
"epoch": 6.0,
|
|
"grad_norm": 0.10400390625,
|
|
"learning_rate": 1.9919548128307954e-05,
|
|
"loss": 0.0003,
|
|
"step": 6
|
|
},
|
|
{
|
|
"epoch": 6.0,
|
|
"eval_accumulated_loss": 1324.264461517334,
|
|
"eval_accuracy": 0.6192982456140351,
|
|
"eval_accuracy_32": 0.6139240506329114,
|
|
"eval_accuracy_33": 0.743421052631579,
|
|
"eval_accuracy_34": 0.6338028169014085,
|
|
"eval_accuracy_35": 0.4491525423728814,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 288.0,
|
|
"eval_correct_gen_preds_32": 64.0,
|
|
"eval_correct_gen_preds_33": 103.0,
|
|
"eval_correct_gen_preds_34": 75.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 353.0,
|
|
"eval_correct_preds_32": 97.0,
|
|
"eval_correct_preds_33": 113.0,
|
|
"eval_correct_preds_34": 90.0,
|
|
"eval_correct_preds_35": 53.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5052631578947369,
|
|
"eval_gen_accuracy_32": 0.4050632911392405,
|
|
"eval_gen_accuracy_33": 0.6776315789473685,
|
|
"eval_gen_accuracy_34": 0.528169014084507,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 2.323270797729492,
|
|
"eval_mdl": 1910.5097714565513,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.2265,
|
|
"eval_samples_per_second": 256.011,
|
|
"eval_steps_per_second": 2.695,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 6
|
|
},
|
|
{
|
|
"epoch": 7.0,
|
|
"grad_norm": 0.0145263671875,
|
|
"learning_rate": 1.9874388886763944e-05,
|
|
"loss": 0.0,
|
|
"step": 7
|
|
},
|
|
{
|
|
"epoch": 7.0,
|
|
"eval_accumulated_loss": 1518.1329746246338,
|
|
"eval_accuracy": 0.6421052631578947,
|
|
"eval_accuracy_32": 0.6392405063291139,
|
|
"eval_accuracy_33": 0.756578947368421,
|
|
"eval_accuracy_34": 0.6690140845070423,
|
|
"eval_accuracy_35": 0.4661016949152542,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 306.0,
|
|
"eval_correct_gen_preds_32": 66.0,
|
|
"eval_correct_gen_preds_33": 108.0,
|
|
"eval_correct_gen_preds_34": 81.0,
|
|
"eval_correct_gen_preds_35": 51.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 366.0,
|
|
"eval_correct_preds_32": 101.0,
|
|
"eval_correct_preds_33": 115.0,
|
|
"eval_correct_preds_34": 95.0,
|
|
"eval_correct_preds_35": 55.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5368421052631579,
|
|
"eval_gen_accuracy_32": 0.4177215189873418,
|
|
"eval_gen_accuracy_33": 0.7105263157894737,
|
|
"eval_gen_accuracy_34": 0.5704225352112676,
|
|
"eval_gen_accuracy_35": 0.4322033898305085,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 2.663391351699829,
|
|
"eval_mdl": 2190.2029139009696,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 3.9086,
|
|
"eval_samples_per_second": 145.834,
|
|
"eval_steps_per_second": 1.535,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 7
|
|
},
|
|
{
|
|
"epoch": 8.0,
|
|
"grad_norm": 0.006195068359375,
|
|
"learning_rate": 1.9819286972627066e-05,
|
|
"loss": 0.0,
|
|
"step": 8
|
|
},
|
|
{
|
|
"epoch": 8.0,
|
|
"eval_accumulated_loss": 1669.1528663635254,
|
|
"eval_accuracy": 0.6403508771929824,
|
|
"eval_accuracy_32": 0.6392405063291139,
|
|
"eval_accuracy_33": 0.743421052631579,
|
|
"eval_accuracy_34": 0.676056338028169,
|
|
"eval_accuracy_35": 0.4661016949152542,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 313.0,
|
|
"eval_correct_gen_preds_32": 67.0,
|
|
"eval_correct_gen_preds_33": 109.0,
|
|
"eval_correct_gen_preds_34": 87.0,
|
|
"eval_correct_gen_preds_35": 50.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 365.0,
|
|
"eval_correct_preds_32": 101.0,
|
|
"eval_correct_preds_33": 113.0,
|
|
"eval_correct_preds_34": 96.0,
|
|
"eval_correct_preds_35": 55.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5491228070175439,
|
|
"eval_gen_accuracy_32": 0.4240506329113924,
|
|
"eval_gen_accuracy_33": 0.7171052631578947,
|
|
"eval_gen_accuracy_34": 0.6126760563380281,
|
|
"eval_gen_accuracy_35": 0.423728813559322,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 2.9283385276794434,
|
|
"eval_mdl": 2408.078562788257,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.0392,
|
|
"eval_samples_per_second": 279.525,
|
|
"eval_steps_per_second": 2.942,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 8
|
|
},
|
|
{
|
|
"epoch": 9.0,
|
|
"grad_norm": 0.0032501220703125,
|
|
"learning_rate": 1.9754297868854075e-05,
|
|
"loss": 0.0,
|
|
"step": 9
|
|
},
|
|
{
|
|
"epoch": 9.0,
|
|
"eval_accumulated_loss": 1789.4745388031006,
|
|
"eval_accuracy": 0.6491228070175439,
|
|
"eval_accuracy_32": 0.6582278481012658,
|
|
"eval_accuracy_33": 0.756578947368421,
|
|
"eval_accuracy_34": 0.6830985915492958,
|
|
"eval_accuracy_35": 0.4576271186440678,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 318.0,
|
|
"eval_correct_gen_preds_32": 70.0,
|
|
"eval_correct_gen_preds_33": 110.0,
|
|
"eval_correct_gen_preds_34": 88.0,
|
|
"eval_correct_gen_preds_35": 50.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 370.0,
|
|
"eval_correct_preds_32": 104.0,
|
|
"eval_correct_preds_33": 115.0,
|
|
"eval_correct_preds_34": 97.0,
|
|
"eval_correct_preds_35": 54.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5578947368421052,
|
|
"eval_gen_accuracy_32": 0.4430379746835443,
|
|
"eval_gen_accuracy_33": 0.7236842105263158,
|
|
"eval_gen_accuracy_34": 0.6197183098591549,
|
|
"eval_gen_accuracy_35": 0.423728813559322,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.1394293308258057,
|
|
"eval_mdl": 2581.6660429282983,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 1.7388,
|
|
"eval_samples_per_second": 327.807,
|
|
"eval_steps_per_second": 3.451,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 9
|
|
},
|
|
{
|
|
"epoch": 10.0,
|
|
"grad_norm": 0.0020751953125,
|
|
"learning_rate": 1.9679487013963566e-05,
|
|
"loss": 0.0,
|
|
"step": 10
|
|
},
|
|
{
|
|
"epoch": 10.0,
|
|
"eval_accumulated_loss": 1878.2606582641602,
|
|
"eval_accuracy": 0.6456140350877193,
|
|
"eval_accuracy_32": 0.6392405063291139,
|
|
"eval_accuracy_33": 0.75,
|
|
"eval_accuracy_34": 0.6901408450704225,
|
|
"eval_accuracy_35": 0.4661016949152542,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 314.0,
|
|
"eval_correct_gen_preds_32": 73.0,
|
|
"eval_correct_gen_preds_33": 109.0,
|
|
"eval_correct_gen_preds_34": 86.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 368.0,
|
|
"eval_correct_preds_32": 101.0,
|
|
"eval_correct_preds_33": 114.0,
|
|
"eval_correct_preds_34": 98.0,
|
|
"eval_correct_preds_35": 55.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5508771929824562,
|
|
"eval_gen_accuracy_32": 0.4620253164556962,
|
|
"eval_gen_accuracy_33": 0.7171052631578947,
|
|
"eval_gen_accuracy_34": 0.6056338028169014,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.295194387435913,
|
|
"eval_mdl": 2709.757337174544,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 4.1173,
|
|
"eval_samples_per_second": 138.44,
|
|
"eval_steps_per_second": 1.457,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 10
|
|
},
|
|
{
|
|
"epoch": 11.0,
|
|
"grad_norm": 0.001251220703125,
|
|
"learning_rate": 1.9594929736144978e-05,
|
|
"loss": 0.0,
|
|
"step": 11
|
|
},
|
|
{
|
|
"epoch": 11.0,
|
|
"eval_accumulated_loss": 1943.7979278564453,
|
|
"eval_accuracy": 0.6421052631578947,
|
|
"eval_accuracy_32": 0.6329113924050633,
|
|
"eval_accuracy_33": 0.75,
|
|
"eval_accuracy_34": 0.6901408450704225,
|
|
"eval_accuracy_35": 0.4576271186440678,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 318.0,
|
|
"eval_correct_gen_preds_32": 74.0,
|
|
"eval_correct_gen_preds_33": 109.0,
|
|
"eval_correct_gen_preds_34": 89.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 366.0,
|
|
"eval_correct_preds_32": 100.0,
|
|
"eval_correct_preds_33": 114.0,
|
|
"eval_correct_preds_34": 98.0,
|
|
"eval_correct_preds_35": 54.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5578947368421052,
|
|
"eval_gen_accuracy_32": 0.46835443037974683,
|
|
"eval_gen_accuracy_33": 0.7171052631578947,
|
|
"eval_gen_accuracy_34": 0.6267605633802817,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.4101719856262207,
|
|
"eval_mdl": 2804.3076310087367,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.459,
|
|
"eval_samples_per_second": 231.805,
|
|
"eval_steps_per_second": 2.44,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 11
|
|
},
|
|
{
|
|
"epoch": 12.0,
|
|
"grad_norm": 0.0009918212890625,
|
|
"learning_rate": 1.9500711177409456e-05,
|
|
"loss": 0.0,
|
|
"step": 12
|
|
},
|
|
{
|
|
"epoch": 12.0,
|
|
"eval_accumulated_loss": 1991.1740913391113,
|
|
"eval_accuracy": 0.6421052631578947,
|
|
"eval_accuracy_32": 0.6329113924050633,
|
|
"eval_accuracy_33": 0.75,
|
|
"eval_accuracy_34": 0.6901408450704225,
|
|
"eval_accuracy_35": 0.4576271186440678,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 320.0,
|
|
"eval_correct_gen_preds_32": 74.0,
|
|
"eval_correct_gen_preds_33": 110.0,
|
|
"eval_correct_gen_preds_34": 90.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 366.0,
|
|
"eval_correct_preds_32": 100.0,
|
|
"eval_correct_preds_33": 114.0,
|
|
"eval_correct_preds_34": 98.0,
|
|
"eval_correct_preds_35": 54.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5614035087719298,
|
|
"eval_gen_accuracy_32": 0.46835443037974683,
|
|
"eval_gen_accuracy_33": 0.7236842105263158,
|
|
"eval_gen_accuracy_34": 0.6338028169014085,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.493288040161133,
|
|
"eval_mdl": 2872.656987121524,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 1.9289,
|
|
"eval_samples_per_second": 295.499,
|
|
"eval_steps_per_second": 3.111,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 12
|
|
},
|
|
{
|
|
"epoch": 13.0,
|
|
"grad_norm": 0.0007476806640625,
|
|
"learning_rate": 1.9396926207859085e-05,
|
|
"loss": 0.0,
|
|
"step": 13
|
|
},
|
|
{
|
|
"epoch": 13.0,
|
|
"eval_accumulated_loss": 2030.5798206329346,
|
|
"eval_accuracy": 0.6385964912280702,
|
|
"eval_accuracy_32": 0.6329113924050633,
|
|
"eval_accuracy_33": 0.743421052631579,
|
|
"eval_accuracy_34": 0.6830985915492958,
|
|
"eval_accuracy_35": 0.4576271186440678,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 318.0,
|
|
"eval_correct_gen_preds_32": 75.0,
|
|
"eval_correct_gen_preds_33": 109.0,
|
|
"eval_correct_gen_preds_34": 88.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 364.0,
|
|
"eval_correct_preds_32": 100.0,
|
|
"eval_correct_preds_33": 113.0,
|
|
"eval_correct_preds_34": 97.0,
|
|
"eval_correct_preds_35": 54.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5578947368421052,
|
|
"eval_gen_accuracy_32": 0.47468354430379744,
|
|
"eval_gen_accuracy_33": 0.7171052631578947,
|
|
"eval_gen_accuracy_34": 0.6197183098591549,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.5624210834503174,
|
|
"eval_mdl": 2929.5074373563357,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 3.2943,
|
|
"eval_samples_per_second": 173.028,
|
|
"eval_steps_per_second": 1.821,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 13
|
|
},
|
|
{
|
|
"epoch": 14.0,
|
|
"grad_norm": 0.00057220458984375,
|
|
"learning_rate": 1.9283679330160726e-05,
|
|
"loss": 0.0,
|
|
"step": 14
|
|
},
|
|
{
|
|
"epoch": 14.0,
|
|
"eval_accumulated_loss": 2057.405590057373,
|
|
"eval_accuracy": 0.6403508771929824,
|
|
"eval_accuracy_32": 0.6329113924050633,
|
|
"eval_accuracy_33": 0.75,
|
|
"eval_accuracy_34": 0.6830985915492958,
|
|
"eval_accuracy_35": 0.4576271186440678,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 318.0,
|
|
"eval_correct_gen_preds_32": 71.0,
|
|
"eval_correct_gen_preds_33": 111.0,
|
|
"eval_correct_gen_preds_34": 90.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 365.0,
|
|
"eval_correct_preds_32": 100.0,
|
|
"eval_correct_preds_33": 114.0,
|
|
"eval_correct_preds_34": 97.0,
|
|
"eval_correct_preds_35": 54.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5578947368421052,
|
|
"eval_gen_accuracy_32": 0.44936708860759494,
|
|
"eval_gen_accuracy_33": 0.7302631578947368,
|
|
"eval_gen_accuracy_34": 0.6338028169014085,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.6094837188720703,
|
|
"eval_mdl": 2968.2088418730036,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.0027,
|
|
"eval_samples_per_second": 284.614,
|
|
"eval_steps_per_second": 2.996,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 14
|
|
},
|
|
{
|
|
"epoch": 15.0,
|
|
"grad_norm": 0.000545501708984375,
|
|
"learning_rate": 1.9161084574320696e-05,
|
|
"loss": 0.0,
|
|
"step": 15
|
|
},
|
|
{
|
|
"epoch": 15.0,
|
|
"eval_accumulated_loss": 2082.8866996765137,
|
|
"eval_accuracy": 0.6456140350877193,
|
|
"eval_accuracy_32": 0.6455696202531646,
|
|
"eval_accuracy_33": 0.743421052631579,
|
|
"eval_accuracy_34": 0.6901408450704225,
|
|
"eval_accuracy_35": 0.4661016949152542,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 319.0,
|
|
"eval_correct_gen_preds_32": 73.0,
|
|
"eval_correct_gen_preds_33": 110.0,
|
|
"eval_correct_gen_preds_34": 90.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 368.0,
|
|
"eval_correct_preds_32": 102.0,
|
|
"eval_correct_preds_33": 113.0,
|
|
"eval_correct_preds_34": 98.0,
|
|
"eval_correct_preds_35": 55.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5596491228070175,
|
|
"eval_gen_accuracy_32": 0.4620253164556962,
|
|
"eval_gen_accuracy_33": 0.7236842105263158,
|
|
"eval_gen_accuracy_34": 0.6338028169014085,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.6541876792907715,
|
|
"eval_mdl": 3004.970312356886,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.6596,
|
|
"eval_samples_per_second": 214.32,
|
|
"eval_steps_per_second": 2.256,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 15
|
|
},
|
|
{
|
|
"epoch": 16.0,
|
|
"grad_norm": 0.0004711151123046875,
|
|
"learning_rate": 1.9029265382866216e-05,
|
|
"loss": 0.0,
|
|
"step": 16
|
|
},
|
|
{
|
|
"epoch": 16.0,
|
|
"eval_accumulated_loss": 2088.4699153900146,
|
|
"eval_accuracy": 0.643859649122807,
|
|
"eval_accuracy_32": 0.6392405063291139,
|
|
"eval_accuracy_33": 0.75,
|
|
"eval_accuracy_34": 0.6901408450704225,
|
|
"eval_accuracy_35": 0.4576271186440678,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 321.0,
|
|
"eval_correct_gen_preds_32": 73.0,
|
|
"eval_correct_gen_preds_33": 111.0,
|
|
"eval_correct_gen_preds_34": 91.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 367.0,
|
|
"eval_correct_preds_32": 101.0,
|
|
"eval_correct_preds_33": 114.0,
|
|
"eval_correct_preds_34": 98.0,
|
|
"eval_correct_preds_35": 54.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5631578947368421,
|
|
"eval_gen_accuracy_32": 0.4620253164556962,
|
|
"eval_gen_accuracy_33": 0.7302631578947368,
|
|
"eval_gen_accuracy_34": 0.6408450704225352,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.66398286819458,
|
|
"eval_mdl": 3013.0251899789673,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 4.8066,
|
|
"eval_samples_per_second": 118.587,
|
|
"eval_steps_per_second": 1.248,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 16
|
|
},
|
|
{
|
|
"epoch": 17.0,
|
|
"grad_norm": 0.00041961669921875,
|
|
"learning_rate": 1.8888354486549238e-05,
|
|
"loss": 0.0,
|
|
"step": 17
|
|
},
|
|
{
|
|
"epoch": 17.0,
|
|
"eval_accumulated_loss": 2108.461343765259,
|
|
"eval_accuracy": 0.6473684210526316,
|
|
"eval_accuracy_32": 0.6518987341772152,
|
|
"eval_accuracy_33": 0.743421052631579,
|
|
"eval_accuracy_34": 0.6901408450704225,
|
|
"eval_accuracy_35": 0.4661016949152542,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 321.0,
|
|
"eval_correct_gen_preds_32": 76.0,
|
|
"eval_correct_gen_preds_33": 109.0,
|
|
"eval_correct_gen_preds_34": 90.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 369.0,
|
|
"eval_correct_preds_32": 103.0,
|
|
"eval_correct_preds_33": 113.0,
|
|
"eval_correct_preds_34": 98.0,
|
|
"eval_correct_preds_35": 55.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5631578947368421,
|
|
"eval_gen_accuracy_32": 0.4810126582278481,
|
|
"eval_gen_accuracy_33": 0.7171052631578947,
|
|
"eval_gen_accuracy_34": 0.6338028169014085,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.6990551948547363,
|
|
"eval_mdl": 3041.8667245562187,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.3278,
|
|
"eval_samples_per_second": 244.866,
|
|
"eval_steps_per_second": 2.578,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 17
|
|
},
|
|
{
|
|
"epoch": 18.0,
|
|
"grad_norm": 0.00040435791015625,
|
|
"learning_rate": 1.873849377069785e-05,
|
|
"loss": 0.0,
|
|
"step": 18
|
|
},
|
|
{
|
|
"epoch": 18.0,
|
|
"eval_accumulated_loss": 2120.72190284729,
|
|
"eval_accuracy": 0.6491228070175439,
|
|
"eval_accuracy_32": 0.6582278481012658,
|
|
"eval_accuracy_33": 0.743421052631579,
|
|
"eval_accuracy_34": 0.6830985915492958,
|
|
"eval_accuracy_35": 0.4745762711864407,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 321.0,
|
|
"eval_correct_gen_preds_32": 77.0,
|
|
"eval_correct_gen_preds_33": 109.0,
|
|
"eval_correct_gen_preds_34": 89.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 370.0,
|
|
"eval_correct_preds_32": 104.0,
|
|
"eval_correct_preds_33": 113.0,
|
|
"eval_correct_preds_34": 97.0,
|
|
"eval_correct_preds_35": 56.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5631578947368421,
|
|
"eval_gen_accuracy_32": 0.4873417721518987,
|
|
"eval_gen_accuracy_33": 0.7171052631578947,
|
|
"eval_gen_accuracy_34": 0.6267605633802817,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.7205653190612793,
|
|
"eval_mdl": 3059.5549723423915,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 5.5641,
|
|
"eval_samples_per_second": 102.443,
|
|
"eval_steps_per_second": 1.078,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 18
|
|
},
|
|
{
|
|
"epoch": 19.0,
|
|
"grad_norm": 0.0003910064697265625,
|
|
"learning_rate": 1.8579834132349773e-05,
|
|
"loss": 0.0,
|
|
"step": 19
|
|
},
|
|
{
|
|
"epoch": 19.0,
|
|
"eval_accumulated_loss": 2125.0127182006836,
|
|
"eval_accuracy": 0.6456140350877193,
|
|
"eval_accuracy_32": 0.6455696202531646,
|
|
"eval_accuracy_33": 0.75,
|
|
"eval_accuracy_34": 0.6901408450704225,
|
|
"eval_accuracy_35": 0.4576271186440678,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 320.0,
|
|
"eval_correct_gen_preds_32": 74.0,
|
|
"eval_correct_gen_preds_33": 110.0,
|
|
"eval_correct_gen_preds_34": 90.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 368.0,
|
|
"eval_correct_preds_32": 102.0,
|
|
"eval_correct_preds_33": 114.0,
|
|
"eval_correct_preds_34": 98.0,
|
|
"eval_correct_preds_35": 54.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5614035087719298,
|
|
"eval_gen_accuracy_32": 0.46835443037974683,
|
|
"eval_gen_accuracy_33": 0.7236842105263158,
|
|
"eval_gen_accuracy_34": 0.6338028169014085,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.728092908859253,
|
|
"eval_mdl": 3065.7453103741027,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 1.8394,
|
|
"eval_samples_per_second": 309.89,
|
|
"eval_steps_per_second": 3.262,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 19
|
|
},
|
|
{
|
|
"epoch": 20.0,
|
|
"grad_norm": 0.0003662109375,
|
|
"learning_rate": 1.8412535328311813e-05,
|
|
"loss": 0.0,
|
|
"step": 20
|
|
},
|
|
{
|
|
"epoch": 20.0,
|
|
"eval_accumulated_loss": 2130.648073196411,
|
|
"eval_accuracy": 0.6473684210526316,
|
|
"eval_accuracy_32": 0.6455696202531646,
|
|
"eval_accuracy_33": 0.75,
|
|
"eval_accuracy_34": 0.6830985915492958,
|
|
"eval_accuracy_35": 0.4745762711864407,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 321.0,
|
|
"eval_correct_gen_preds_32": 77.0,
|
|
"eval_correct_gen_preds_33": 109.0,
|
|
"eval_correct_gen_preds_34": 90.0,
|
|
"eval_correct_gen_preds_35": 45.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 369.0,
|
|
"eval_correct_preds_32": 102.0,
|
|
"eval_correct_preds_33": 114.0,
|
|
"eval_correct_preds_34": 97.0,
|
|
"eval_correct_preds_35": 56.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5631578947368421,
|
|
"eval_gen_accuracy_32": 0.4873417721518987,
|
|
"eval_gen_accuracy_33": 0.7171052631578947,
|
|
"eval_gen_accuracy_34": 0.6338028169014085,
|
|
"eval_gen_accuracy_35": 0.3813559322033898,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.7379796504974365,
|
|
"eval_mdl": 3073.8754090800876,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.5005,
|
|
"eval_samples_per_second": 227.954,
|
|
"eval_steps_per_second": 2.4,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 20
|
|
},
|
|
{
|
|
"epoch": 21.0,
|
|
"grad_norm": 0.0003643035888671875,
|
|
"learning_rate": 1.8236765814298328e-05,
|
|
"loss": 0.0,
|
|
"step": 21
|
|
},
|
|
{
|
|
"epoch": 21.0,
|
|
"eval_accumulated_loss": 2130.965400695801,
|
|
"eval_accuracy": 0.6508771929824562,
|
|
"eval_accuracy_32": 0.6518987341772152,
|
|
"eval_accuracy_33": 0.743421052631579,
|
|
"eval_accuracy_34": 0.6971830985915493,
|
|
"eval_accuracy_35": 0.4745762711864407,
|
|
"eval_accuracy_36": 0.0,
|
|
"eval_correct_gen_preds": 326.0,
|
|
"eval_correct_gen_preds_32": 79.0,
|
|
"eval_correct_gen_preds_33": 110.0,
|
|
"eval_correct_gen_preds_34": 91.0,
|
|
"eval_correct_gen_preds_35": 46.0,
|
|
"eval_correct_gen_preds_36": 0.0,
|
|
"eval_correct_preds": 371.0,
|
|
"eval_correct_preds_32": 103.0,
|
|
"eval_correct_preds_33": 113.0,
|
|
"eval_correct_preds_34": 99.0,
|
|
"eval_correct_preds_35": 56.0,
|
|
"eval_correct_preds_36": 0.0,
|
|
"eval_gen_accuracy": 0.5719298245614035,
|
|
"eval_gen_accuracy_32": 0.5,
|
|
"eval_gen_accuracy_33": 0.7236842105263158,
|
|
"eval_gen_accuracy_34": 0.6408450704225352,
|
|
"eval_gen_accuracy_35": 0.3898305084745763,
|
|
"eval_gen_accuracy_36": 0.0,
|
|
"eval_loss": 3.7385361194610596,
|
|
"eval_mdl": 3074.3332158897947,
|
|
"eval_model_preparation_time": 0.0069,
|
|
"eval_runtime": 2.367,
|
|
"eval_samples_per_second": 240.814,
|
|
"eval_steps_per_second": 2.535,
|
|
"eval_total_labels_32": 158.0,
|
|
"eval_total_labels_33": 152.0,
|
|
"eval_total_labels_34": 142.0,
|
|
"eval_total_labels_35": 118.0,
|
|
"eval_total_labels_36": 0.0,
|
|
"eval_total_preds": 570.0,
|
|
"step": 21
|
|
}
|
|
],
|
|
"logging_steps": 1,
|
|
"max_steps": 100,
|
|
"num_input_tokens_seen": 0,
|
|
"num_train_epochs": 100,
|
|
"save_steps": 500,
|
|
"stateful_callbacks": {
|
|
"EarlyStoppingCallback": {
|
|
"args": {
|
|
"early_stopping_patience": 30,
|
|
"early_stopping_threshold": 0.0
|
|
},
|
|
"attributes": {
|
|
"early_stopping_patience_counter": 0
|
|
}
|
|
},
|
|
"TrainerControl": {
|
|
"args": {
|
|
"should_epoch_stop": false,
|
|
"should_evaluate": false,
|
|
"should_log": false,
|
|
"should_save": true,
|
|
"should_training_stop": false
|
|
},
|
|
"attributes": {}
|
|
}
|
|
},
|
|
"total_flos": 147139692134400.0,
|
|
"train_batch_size": 64,
|
|
"trial_name": null,
|
|
"trial_params": null
|
|
}
|