Added new results

models/eval/eval_results.json

@@ -1 +1 @@
-{"eval_loss": 0.6932736039161682, "eval_accuracy": 0.5, "eval_f1_macro": 0.4620003056568347, "eval_f1_weighted": 0.46200030565683464, "eval_precision": 0.5, "eval_recall": 0.5, "eval_auroc": 0.5, "eval_runtime": 2.6124, "eval_samples_per_second": 254.941, "eval_steps_per_second": 8.039}
+{"eval_loss": 0.6934633851051331, "eval_accuracy": 0.48375451263537905, "eval_f1_macro": 0.4440795475278234, "eval_f1_weighted": 0.4435433993506942, "eval_precision": 0.478506455399061, "eval_recall": 0.48472526326764676, "eval_auroc": 0.48472526326764676, "eval_runtime": 1.3678, "eval_samples_per_second": 202.514, "eval_steps_per_second": 6.58}

models/semi-supervised/finetune_nested/eval_results.json

@@ -1 +1 @@
-{"0.25": {"eval_loss": 0.4045407474040985, "eval_accuracy": 0.8948948948948949, "eval_f1_macro": 0.8948863636363636, "eval_f1_weighted": 0.8948863636363638, "eval_precision": 0.8950231387513192, "eval_recall": 0.8948948948948949, "eval_auroc": 0.8948948948948949, "eval_runtime": 2.3628, "eval_samples_per_second": 281.865, "eval_steps_per_second": 2.963, "epoch": 25.0}, "0.5": {"eval_loss": 0.44522979855537415, "eval_accuracy": 0.8933933933933934, "eval_f1_macro": 0.8933931530475364, "eval_f1_weighted": 0.8933931530475363, "eval_precision": 0.8933969410576438, "eval_recall": 0.8933933933933934, "eval_auroc": 0.8933933933933934, "eval_runtime": 2.3526, "eval_samples_per_second": 283.09, "eval_steps_per_second": 2.975, "epoch": 25.0}, "0.75": {"eval_loss": 0.5143420696258545, "eval_accuracy": 0.8828828828828829, "eval_f1_macro": 0.8828733766233766, "eval_f1_weighted": 0.8828733766233766, "eval_precision": 0.8830072257854997, "eval_recall": 0.8828828828828829, "eval_auroc": 0.8828828828828829, "eval_runtime": 2.353, "eval_samples_per_second": 283.043, "eval_steps_per_second": 2.975, "epoch": 25.0}, "1.0": {"eval_loss": 0.5319307446479797, "eval_accuracy": 0.8828828828828829, "eval_f1_macro": 0.8828818267080297, "eval_f1_weighted": 0.8828818267080298, "eval_precision": 0.8828966947738648, "eval_recall": 0.882882882882883, "eval_auroc": 0.8828828828828829, "eval_runtime": 2.3732, "eval_samples_per_second": 280.639, "eval_steps_per_second": 2.95, "epoch": 25.0}}
+{"0.25": {"eval_loss": 0.32715415954589844, "eval_accuracy": 0.855595667870036, "eval_f1_macro": 0.8555486024196912, "eval_f1_weighted": 0.8555391893296223, "eval_precision": 0.8562291144527987, "eval_recall": 0.855671984151809, "eval_auroc": 0.855671984151809, "eval_runtime": 0.9815, "eval_samples_per_second": 282.211, "eval_steps_per_second": 3.056, "epoch": 25.0}, "0.5": {"eval_loss": 0.3914715349674225, "eval_accuracy": 0.8592057761732852, "eval_f1_macro": 0.8590220412637513, "eval_f1_weighted": 0.8590404147547047, "eval_precision": 0.860779384035198, "eval_recall": 0.8590866437284954, "eval_auroc": 0.8590866437284955, "eval_runtime": 0.9827, "eval_samples_per_second": 281.872, "eval_steps_per_second": 3.053, "epoch": 25.0}, "0.75": {"eval_loss": 0.5000463128089905, "eval_accuracy": 0.8411552346570397, "eval_f1_macro": 0.840987370838117, "eval_f1_weighted": 0.8410060223735529, "eval_precision": 0.8423338566195709, "eval_recall": 0.8410489000104264, "eval_auroc": 0.8410489000104265, "eval_runtime": 0.9822, "eval_samples_per_second": 282.008, "eval_steps_per_second": 3.054, "epoch": 25.0}, "1.0": {"eval_loss": 0.49611490964889526, "eval_accuracy": 0.8772563176895307, "eval_f1_macro": 0.877062447786132, "eval_f1_weighted": 0.8770800723228045, "eval_precision": 0.8792994966442953, "eval_recall": 0.8771243874465644, "eval_auroc": 0.8771243874465645, "eval_runtime": 0.9822, "eval_samples_per_second": 282.031, "eval_steps_per_second": 3.054, "epoch": 25.0}}

models/supervised/finetune_nested/eval_results.json

@@ -1 +1 @@
-{"0.25": {"eval_loss": 0.691838800907135, "eval_accuracy": 0.5675675675675675, "eval_f1_macro": 0.5675675675675675, "eval_f1_weighted": 0.5675675675675675, "eval_precision": 0.5675675675675675, "eval_recall": 0.5675675675675675, "eval_auroc": 0.5675675675675675, "eval_runtime": 2.3948, "eval_samples_per_second": 278.108, "eval_steps_per_second": 2.923, "epoch": 25.0}, "0.5": {"eval_loss": 0.6798407435417175, "eval_accuracy": 0.6816816816816816, "eval_f1_macro": 0.6765360824742268, "eval_f1_weighted": 0.6765360824742268, "eval_precision": 0.6940279102019589, "eval_recall": 0.6816816816816818, "eval_auroc": 0.6816816816816816, "eval_runtime": 2.3669, "eval_samples_per_second": 281.381, "eval_steps_per_second": 2.957, "epoch": 25.0}, "0.75": {"eval_loss": 0.3387678563594818, "eval_accuracy": 0.8843843843843844, "eval_f1_macro": 0.8843778676124103, "eval_f1_weighted": 0.8843778676124106, "eval_precision": 0.8844710636455477, "eval_recall": 0.8843843843843844, "eval_auroc": 0.8843843843843844, "eval_runtime": 2.3506, "eval_samples_per_second": 283.327, "eval_steps_per_second": 2.978, "epoch": 25.0}, "1.0": {"eval_loss": 0.2984059453010559, "eval_accuracy": 0.8993993993993994, "eval_f1_macro": 0.8992992789682137, "eval_f1_weighted": 0.8992992789682136, "eval_precision": 0.9009941329856584, "eval_recall": 0.8993993993993994, "eval_auroc": 0.8993993993993994, "eval_runtime": 2.3504, "eval_samples_per_second": 283.362, "eval_steps_per_second": 2.978, "epoch": 25.0}}
+{"0.25": {"eval_loss": 0.6932923793792725, "eval_accuracy": 0.49097472924187724, "eval_f1_macro": 0.4247271441827582, "eval_f1_weighted": 0.42402238263957603, "eval_precision": 0.48541747951619196, "eval_recall": 0.49220623501199046, "eval_auroc": 0.4922062350119904, "eval_runtime": 0.9797, "eval_samples_per_second": 282.736, "eval_steps_per_second": 3.062, "epoch": 25.0}, "0.5": {"eval_loss": 0.691850483417511, "eval_accuracy": 0.5451263537906137, "eval_f1_macro": 0.5025937749401437, "eval_f1_weighted": 0.5031188685061988, "eval_precision": 0.5665643205794363, "eval_recall": 0.5440777812532582, "eval_auroc": 0.5440777812532582, "eval_runtime": 0.9804, "eval_samples_per_second": 282.536, "eval_steps_per_second": 3.06, "epoch": 25.0}, "0.75": {"eval_loss": 0.6915454268455505, "eval_accuracy": 0.5451263537906137, "eval_f1_macro": 0.48733842538190364, "eval_f1_weighted": 0.4879598009561908, "eval_precision": 0.5792866553736119, "eval_recall": 0.5439213846314253, "eval_auroc": 0.5439213846314253, "eval_runtime": 0.9814, "eval_samples_per_second": 282.253, "eval_steps_per_second": 3.057, "epoch": 25.0}, "1.0": {"eval_loss": 0.6900911927223206, "eval_accuracy": 0.5992779783393501, "eval_f1_macro": 0.5907439204568142, "eval_f1_weighted": 0.5909572719038776, "eval_precision": 0.6075076608784473, "eval_recall": 0.5987644666875195, "eval_auroc": 0.5987644666875196, "eval_runtime": 0.982, "eval_samples_per_second": 282.09, "eval_steps_per_second": 3.055, "epoch": 25.0}}

models/supervised/transfer_nested/eval_results.json

@@ -1 +1 @@
-{"0.25": {"eval_loss": 0.6932699680328369, "eval_accuracy": 0.496996996996997, "eval_f1_macro": 0.4623790648080097, "eval_f1_weighted": 0.4623790648080097, "eval_precision": 0.49595520357594014, "eval_recall": 0.49699699699699695, "eval_auroc": 0.49699699699699695, "eval_runtime": 2.3519, "eval_samples_per_second": 283.169, "eval_steps_per_second": 2.976, "epoch": 25.0}, "0.5": {"eval_loss": 0.6928256154060364, "eval_accuracy": 0.524024024024024, "eval_f1_macro": 0.48958361035425635, "eval_f1_weighted": 0.48958361035425635, "eval_precision": 0.532905138339921, "eval_recall": 0.524024024024024, "eval_auroc": 0.524024024024024, "eval_runtime": 2.3686, "eval_samples_per_second": 281.18, "eval_steps_per_second": 2.955, "epoch": 25.0}, "0.75": {"eval_loss": 0.6927742958068848, "eval_accuracy": 0.521021021021021, "eval_f1_macro": 0.4888841308066313, "eval_f1_weighted": 0.4888841308066313, "eval_precision": 0.5280843373493975, "eval_recall": 0.5210210210210211, "eval_auroc": 0.521021021021021, "eval_runtime": 2.35, "eval_samples_per_second": 283.403, "eval_steps_per_second": 2.979, "epoch": 25.0}, "1.0": {"eval_loss": 0.6926871538162231, "eval_accuracy": 0.5195195195195195, "eval_f1_macro": 0.48930296756383707, "eval_f1_weighted": 0.48930296756383707, "eval_precision": 0.5255715045188729, "eval_recall": 0.5195195195195195, "eval_auroc": 0.5195195195195196, "eval_runtime": 2.3534, "eval_samples_per_second": 282.994, "eval_steps_per_second": 2.974, "epoch": 25.0}}
+{"0.25": {"eval_loss": 0.6934651136398315, "eval_accuracy": 0.48375451263537905, "eval_f1_macro": 0.4440795475278234, "eval_f1_weighted": 0.4435433993506942, "eval_precision": 0.478506455399061, "eval_recall": 0.48472526326764676, "eval_auroc": 0.48472526326764676, "eval_runtime": 0.9834, "eval_samples_per_second": 281.682, "eval_steps_per_second": 3.051, "epoch": 25.0}, "0.5": {"eval_loss": 0.6924956440925598, "eval_accuracy": 0.5342960288808665, "eval_f1_macro": 0.4941892561398542, "eval_f1_weighted": 0.4947034455339697, "eval_precision": 0.5484593199757134, "eval_recall": 0.5332864143467835, "eval_auroc": 0.5332864143467835, "eval_runtime": 0.9856, "eval_samples_per_second": 281.049, "eval_steps_per_second": 3.044, "epoch": 25.0}, "0.75": {"eval_loss": 0.6925008893013, "eval_accuracy": 0.5342960288808665, "eval_f1_macro": 0.4941892561398542, "eval_f1_weighted": 0.4947034455339697, "eval_precision": 0.5484593199757134, "eval_recall": 0.5332864143467835, "eval_auroc": 0.5332864143467835, "eval_runtime": 0.9798, "eval_samples_per_second": 282.72, "eval_steps_per_second": 3.062, "epoch": 25.0}, "1.0": {"eval_loss": 0.6925026774406433, "eval_accuracy": 0.5342960288808665, "eval_f1_macro": 0.4941892561398542, "eval_f1_weighted": 0.4947034455339697, "eval_precision": 0.5484593199757134, "eval_recall": 0.5332864143467835, "eval_auroc": 0.5332864143467835, "eval_runtime": 0.9818, "eval_samples_per_second": 282.129, "eval_steps_per_second": 3.056, "epoch": 25.0}}