{ "best_metric": 0.7475216864388343, "best_model_checkpoint": "checkpoint/reproduce/unsup-bert-base-uncased/checkpoint-20000", "epoch": 3.0, "global_step": 46875, "is_hyper_param_search": false, "is_local_process_zero": true, "is_world_process_zero": true, "log_history": [ { "epoch": 0.04, "learning_rate": 2.2182723071453966e-06, "loss": 0.0182, "step": 625 }, { "epoch": 0.04, "eval_avg_sts": 0.5761493170311456, "eval_avg_test": 0.5719267013539271, "eval_avg_train_and_dev": 0.5955653275065955, "eval_runtime": 20.9393, "eval_sickrelatedness_spearman": 0.614035610432408, "eval_sickrelatedness_test_spearman": 0.6110088391931596, "eval_sickrelatedness_train_and_dev_spearman": 0.6170011765653474, "eval_sts12_spearman": 0.3883511001773007, "eval_sts13_spearman": 0.6409334821384267, "eval_sts14_spearman": 0.5233435331267315, "eval_sts15_spearman": 0.6304876176600429, "eval_sts16_spearman": 0.6668361049295604, "eval_stsbenchmark_spearman": 0.569057770753549, "eval_stsbenchmark_test_spearman": 0.5425262322522679, "eval_stsbenchmark_train_and_dev_spearman": 0.5741294784478437, "step": 625 }, { "epoch": 0.08, "learning_rate": 4.440099537859936e-06, "loss": 0.0009, "step": 1250 }, { "epoch": 0.08, "eval_avg_sts": 0.6176613225678802, "eval_avg_test": 0.6146526994706545, "eval_avg_train_and_dev": 0.6354918146843727, "eval_runtime": 22.0658, "eval_sickrelatedness_spearman": 0.6481782560461822, "eval_sickrelatedness_test_spearman": 0.6454675815413363, "eval_sickrelatedness_train_and_dev_spearman": 0.6510278427067593, "eval_sts12_spearman": 0.4492677971681109, "eval_sts13_spearman": 0.6749925740490339, "eval_sts14_spearman": 0.5645147644908317, "eval_sts15_spearman": 0.6736146568642941, "eval_sts16_spearman": 0.6965736684860125, "eval_stsbenchmark_spearman": 0.6164875408706965, "eval_stsbenchmark_test_spearman": 0.5981378536949618, "eval_stsbenchmark_train_and_dev_spearman": 0.6199557866619861, "step": 1250 }, { "epoch": 0.12, "learning_rate": 6.661926768574476e-06, "loss": 0.0003, "step": 1875 }, { "epoch": 0.12, "eval_avg_sts": 0.6494555691959818, "eval_avg_test": 0.6471419605134568, "eval_avg_train_and_dev": 0.6641694696137254, "eval_runtime": 21.9114, "eval_sickrelatedness_spearman": 0.6689216651276976, "eval_sickrelatedness_test_spearman": 0.6660745763942842, "eval_sickrelatedness_train_and_dev_spearman": 0.6717733112444599, "eval_sts12_spearman": 0.4998164334853574, "eval_sts13_spearman": 0.7060812594730672, "eval_sts14_spearman": 0.5974341417069697, "eval_sts15_spearman": 0.7041580385169013, "eval_sts16_spearman": 0.7158255254242747, "eval_stsbenchmark_spearman": 0.6539519206376052, "eval_stsbenchmark_test_spearman": 0.6406037485933435, "eval_stsbenchmark_train_and_dev_spearman": 0.6565656279829908, "step": 1875 }, { "epoch": 0.16, "learning_rate": 8.883753999289016e-06, "loss": 0.0003, "step": 2500 }, { "epoch": 0.16, "eval_avg_sts": 0.6720172723077595, "eval_avg_test": 0.6700901794645714, "eval_avg_train_and_dev": 0.6813887988919534, "eval_runtime": 21.6305, "eval_sickrelatedness_spearman": 0.6752369370120973, "eval_sickrelatedness_test_spearman": 0.6720545832463273, "eval_sickrelatedness_train_and_dev_spearman": 0.6784163347898827, "eval_sts12_spearman": 0.5357838277332643, "eval_sts13_spearman": 0.724945557702781, "eval_sts14_spearman": 0.6231065986339928, "eval_sts15_spearman": 0.7325580598726732, "eval_sts16_spearman": 0.7302626467999829, "eval_stsbenchmark_spearman": 0.6822272783995248, "eval_stsbenchmark_test_spearman": 0.6719199822629783, "eval_stsbenchmark_train_and_dev_spearman": 0.6843612629940239, "step": 2500 }, { "epoch": 0.2, "learning_rate": 9.929417638781718e-06, "loss": 0.0002, "step": 3125 }, { "epoch": 0.2, "eval_avg_sts": 0.6909069613326249, "eval_avg_test": 0.6892152885857711, "eval_avg_train_and_dev": 0.6981092005408447, "eval_runtime": 21.3949, "eval_sickrelatedness_spearman": 0.684142432010355, "eval_sickrelatedness_test_spearman": 0.6810570206913322, "eval_sickrelatedness_train_and_dev_spearman": 0.6872006477750381, "eval_sts12_spearman": 0.571210204194906, "eval_sts13_spearman": 0.744024983013372, "eval_sts14_spearman": 0.6427295960479041, "eval_sts15_spearman": 0.75072362281639, "eval_sts16_spearman": 0.7362828177459125, "eval_stsbenchmark_spearman": 0.7072350734995339, "eval_stsbenchmark_test_spearman": 0.6984787755905806, "eval_stsbenchmark_train_and_dev_spearman": 0.7090177533066512, "step": 3125 }, { "epoch": 0.24, "learning_rate": 9.787572057555263e-06, "loss": 0.0001, "step": 3750 }, { "epoch": 0.24, "eval_avg_sts": 0.7110034740131804, "eval_avg_test": 0.7090073777235183, "eval_avg_train_and_dev": 0.7152562951128082, "eval_runtime": 21.4902, "eval_sickrelatedness_spearman": 0.6924481449933352, "eval_sickrelatedness_test_spearman": 0.6896546207158979, "eval_sickrelatedness_train_and_dev_spearman": 0.6953282368083006, "eval_sts12_spearman": 0.5925559956254365, "eval_sts13_spearman": 0.7670195193741965, "eval_sts14_spearman": 0.6676224547832722, "eval_sts15_spearman": 0.7714028016335752, "eval_sts16_spearman": 0.7530593474380043, "eval_stsbenchmark_spearman": 0.7329160542444434, "eval_stsbenchmark_test_spearman": 0.7217369044942459, "eval_stsbenchmark_train_and_dev_spearman": 0.7351843534173158, "step": 3750 }, { "epoch": 0.28, "learning_rate": 9.645726476328811e-06, "loss": 0.0002, "step": 4375 }, { "epoch": 0.28, "eval_avg_sts": 0.7194580287637145, "eval_avg_test": 0.717033045368064, "eval_avg_train_and_dev": 0.72525371399646, "eval_runtime": 21.832, "eval_sickrelatedness_spearman": 0.69964732104727, "eval_sickrelatedness_test_spearman": 0.6974584549846506, "eval_sickrelatedness_train_and_dev_spearman": 0.7019540331749327, "eval_sts12_spearman": 0.6087565607405551, "eval_sts13_spearman": 0.775181610329269, "eval_sts14_spearman": 0.6807344815935793, "eval_sts15_spearman": 0.7714449656594323, "eval_sts16_spearman": 0.7549091580553093, "eval_stsbenchmark_spearman": 0.7455321039205858, "eval_stsbenchmark_test_spearman": 0.7307460862136526, "eval_stsbenchmark_train_and_dev_spearman": 0.7485533948179873, "step": 4375 }, { "epoch": 0.32, "learning_rate": 9.503880895102356e-06, "loss": 0.0002, "step": 5000 }, { "epoch": 0.32, "eval_avg_sts": 0.7357926871862974, "eval_avg_test": 0.7337295361557176, "eval_avg_train_and_dev": 0.7387626368976384, "eval_runtime": 21.3496, "eval_sickrelatedness_spearman": 0.7081277931755735, "eval_sickrelatedness_test_spearman": 0.7070731150129818, "eval_sickrelatedness_train_and_dev_spearman": 0.7093962696914289, "eval_sts12_spearman": 0.6344626415135579, "eval_sts13_spearman": 0.7877078434493416, "eval_sts14_spearman": 0.6986755801629566, "eval_sts15_spearman": 0.7910799783582211, "eval_sts16_spearman": 0.7651327511753051, "eval_stsbenchmark_spearman": 0.7653622224691259, "eval_stsbenchmark_test_spearman": 0.7519748434176595, "eval_stsbenchmark_train_and_dev_spearman": 0.768129004103848, "step": 5000 }, { "epoch": 0.36, "learning_rate": 9.362035313875902e-06, "loss": 0.0001, "step": 5625 }, { "epoch": 0.36, "eval_avg_sts": 0.7418773860073935, "eval_avg_test": 0.7395590037062557, "eval_avg_train_and_dev": 0.743014478903171, "eval_runtime": 22.6753, "eval_sickrelatedness_spearman": 0.7091920474244533, "eval_sickrelatedness_test_spearman": 0.7076497343346897, "eval_sickrelatedness_train_and_dev_spearman": 0.7108735600549281, "eval_sts12_spearman": 0.6459740297105186, "eval_sts13_spearman": 0.7946896206600376, "eval_sts14_spearman": 0.706190273571107, "eval_sts15_spearman": 0.7965399044791899, "eval_sts16_spearman": 0.7684287092404692, "eval_stsbenchmark_spearman": 0.7721271169659789, "eval_stsbenchmark_test_spearman": 0.7574407539477783, "eval_stsbenchmark_train_and_dev_spearman": 0.7751553977514137, "step": 5625 }, { "epoch": 0.4, "learning_rate": 9.220189732649449e-06, "loss": 0.0001, "step": 6250 }, { "epoch": 0.4, "eval_avg_sts": 0.7308385327663487, "eval_avg_test": 0.7283863218628899, "eval_avg_train_and_dev": 0.7345345112957317, "eval_runtime": 21.8169, "eval_sickrelatedness_spearman": 0.7046853908818935, "eval_sickrelatedness_test_spearman": 0.7037265532974459, "eval_sickrelatedness_train_and_dev_spearman": 0.7057337293264881, "eval_sts12_spearman": 0.621875598643423, "eval_sts13_spearman": 0.7860303924520814, "eval_sts14_spearman": 0.694793783643542, "eval_sts15_spearman": 0.786890544888634, "eval_sts16_spearman": 0.761570793834228, "eval_stsbenchmark_spearman": 0.7600232250206383, "eval_stsbenchmark_test_spearman": 0.7438165862808751, "eval_stsbenchmark_train_and_dev_spearman": 0.7633352932649752, "step": 6250 }, { "epoch": 0.44, "learning_rate": 9.078344151422995e-06, "loss": 0.0002, "step": 6875 }, { "epoch": 0.44, "eval_avg_sts": 0.7309661390558988, "eval_avg_test": 0.7286727854196379, "eval_avg_train_and_dev": 0.7359208652832265, "eval_runtime": 21.1351, "eval_sickrelatedness_spearman": 0.7048537796254629, "eval_sickrelatedness_test_spearman": 0.702412783558349, "eval_sickrelatedness_train_and_dev_spearman": 0.7073541509053075, "eval_sts12_spearman": 0.6186911062602717, "eval_sts13_spearman": 0.7856611011058849, "eval_sts14_spearman": 0.6972845097860244, "eval_sts15_spearman": 0.7842731788758606, "eval_sts16_spearman": 0.7642482925335029, "eval_stsbenchmark_spearman": 0.761751005204284, "eval_stsbenchmark_test_spearman": 0.7481385258175712, "eval_stsbenchmark_train_and_dev_spearman": 0.7644875796611457, "step": 6875 }, { "epoch": 0.48, "learning_rate": 8.936498570196542e-06, "loss": 0.0001, "step": 7500 }, { "epoch": 0.48, "eval_avg_sts": 0.7398140587015002, "eval_avg_test": 0.7375789283159258, "eval_avg_train_and_dev": 0.7436463146851804, "eval_runtime": 21.5092, "eval_sickrelatedness_spearman": 0.7104350424990357, "eval_sickrelatedness_test_spearman": 0.707645634808155, "eval_sickrelatedness_train_and_dev_spearman": 0.7133276839990611, "eval_sts12_spearman": 0.6311869476757046, "eval_sts13_spearman": 0.7943880897002324, "eval_sts14_spearman": 0.7054140851208188, "eval_sts15_spearman": 0.7965258267671331, "eval_sts16_spearman": 0.7693214228482549, "eval_stsbenchmark_spearman": 0.7714269962993212, "eval_stsbenchmark_test_spearman": 0.7585704912911815, "eval_stsbenchmark_train_and_dev_spearman": 0.7739649453712996, "step": 7500 }, { "epoch": 0.52, "learning_rate": 8.794652988970088e-06, "loss": 0.0001, "step": 8125 }, { "epoch": 0.52, "eval_avg_sts": 0.7373193093280432, "eval_avg_test": 0.7349110835074212, "eval_avg_train_and_dev": 0.74065035952663, "eval_runtime": 21.2362, "eval_sickrelatedness_spearman": 0.7089799957481375, "eval_sickrelatedness_test_spearman": 0.7065719679366725, "eval_sickrelatedness_train_and_dev_spearman": 0.711364735604823, "eval_sts12_spearman": 0.62980351920814, "eval_sts13_spearman": 0.7927297021563888, "eval_sts14_spearman": 0.7002451810649419, "eval_sts15_spearman": 0.7943326132694016, "eval_sts16_spearman": 0.7680353970909194, "eval_stsbenchmark_spearman": 0.767108756758373, "eval_stsbenchmark_test_spearman": 0.7526592038254838, "eval_stsbenchmark_train_and_dev_spearman": 0.769935983448437, "step": 8125 }, { "epoch": 0.56, "learning_rate": 8.652807407743635e-06, "loss": 0.0001, "step": 8750 }, { "epoch": 0.56, "eval_avg_sts": 0.7363613158654552, "eval_avg_test": 0.7344457311029845, "eval_avg_train_and_dev": 0.7385101274896235, "eval_runtime": 21.007, "eval_sickrelatedness_spearman": 0.7079709220981268, "eval_sickrelatedness_test_spearman": 0.7060740065988659, "eval_sickrelatedness_train_and_dev_spearman": 0.7098373581622875, "eval_sts12_spearman": 0.6245604819610331, "eval_sts13_spearman": 0.7940661258079552, "eval_sts14_spearman": 0.6995530567749194, "eval_sts15_spearman": 0.7934026178932524, "eval_sts16_spearman": 0.7701123683553529, "eval_stsbenchmark_spearman": 0.764863638167547, "eval_stsbenchmark_test_spearman": 0.7533514603295133, "eval_stsbenchmark_train_and_dev_spearman": 0.7671828968169594, "step": 8750 }, { "epoch": 0.6, "learning_rate": 8.510961826517181e-06, "loss": 0.0001, "step": 9375 }, { "epoch": 0.6, "eval_avg_sts": 0.7318749663720026, "eval_avg_test": 0.7298439077988678, "eval_avg_train_and_dev": 0.7324391542106021, "eval_runtime": 21.8218, "eval_sickrelatedness_spearman": 0.7058228056610777, "eval_sickrelatedness_test_spearman": 0.7039121780620011, "eval_sickrelatedness_train_and_dev_spearman": 0.7077349771782333, "eval_sts12_spearman": 0.618700449208015, "eval_sts13_spearman": 0.7954576133369399, "eval_sts14_spearman": 0.6962032852506622, "eval_sts15_spearman": 0.7872116068915151, "eval_sts16_spearman": 0.765104808041975, "eval_stsbenchmark_spearman": 0.7546241962138335, "eval_stsbenchmark_test_spearman": 0.7423174138009657, "eval_stsbenchmark_train_and_dev_spearman": 0.7571433312429708, "step": 9375 }, { "epoch": 0.64, "learning_rate": 8.369116245290728e-06, "loss": 0.0001, "step": 10000 }, { "epoch": 0.64, "eval_avg_sts": 0.7409280134068449, "eval_avg_test": 0.7391325218596417, "eval_avg_train_and_dev": 0.7415103794141908, "eval_runtime": 21.4143, "eval_sickrelatedness_spearman": 0.7104368210609439, "eval_sickrelatedness_test_spearman": 0.7079528143535209, "eval_sickrelatedness_train_and_dev_spearman": 0.7129369206822436, "eval_sts12_spearman": 0.6369900901886509, "eval_sts13_spearman": 0.8003649874497492, "eval_sts14_spearman": 0.7048130149249648, "eval_sts15_spearman": 0.7956828867682607, "eval_sts16_spearman": 0.7700745630771816, "eval_stsbenchmark_spearman": 0.7681337303781637, "eval_stsbenchmark_test_spearman": 0.7580492962551647, "eval_stsbenchmark_train_and_dev_spearman": 0.770083838146138, "step": 10000 }, { "epoch": 0.68, "learning_rate": 8.227497616994236e-06, "loss": 0.0003, "step": 10625 }, { "epoch": 0.68, "eval_avg_sts": 0.7359543470984595, "eval_avg_test": 0.733931880207311, "eval_avg_train_and_dev": 0.7367064905162184, "eval_runtime": 21.2567, "eval_sickrelatedness_spearman": 0.7081098389229429, "eval_sickrelatedness_test_spearman": 0.7069070856977242, "eval_sickrelatedness_train_and_dev_spearman": 0.7093919257525952, "eval_sts12_spearman": 0.6171012791250989, "eval_sts13_spearman": 0.8001733717764794, "eval_sts14_spearman": 0.7010514936978448, "eval_sts15_spearman": 0.7914208067428021, "eval_sts16_spearman": 0.7723885383628244, "eval_stsbenchmark_spearman": 0.7614351010612246, "eval_stsbenchmark_test_spearman": 0.7484805860484034, "eval_stsbenchmark_train_and_dev_spearman": 0.7640210552798415, "step": 10625 }, { "epoch": 0.72, "learning_rate": 8.085652035767782e-06, "loss": 0.0002, "step": 11250 }, { "epoch": 0.72, "eval_avg_sts": 0.7331832920502, "eval_avg_test": 0.7296440457200707, "eval_avg_train_and_dev": 0.7365755808733788, "eval_runtime": 22.0032, "eval_sickrelatedness_spearman": 0.7101591391042099, "eval_sickrelatedness_test_spearman": 0.7080075712587889, "eval_sickrelatedness_train_and_dev_spearman": 0.7123952349576758, "eval_sts12_spearman": 0.6071610926796485, "eval_sts13_spearman": 0.7986426077282267, "eval_sts14_spearman": 0.7048385507770369, "eval_sts15_spearman": 0.790486863635161, "eval_sts16_spearman": 0.7648214181893283, "eval_stsbenchmark_spearman": 0.7561733722377888, "eval_stsbenchmark_test_spearman": 0.7335502157723038, "eval_stsbenchmark_train_and_dev_spearman": 0.7607559267890819, "step": 11250 }, { "epoch": 0.76, "learning_rate": 7.943806454541329e-06, "loss": 0.0001, "step": 11875 }, { "epoch": 0.76, "eval_avg_sts": 0.7284741669636894, "eval_avg_test": 0.7249847364931884, "eval_avg_train_and_dev": 0.7313381333368812, "eval_runtime": 21.7362, "eval_sickrelatedness_spearman": 0.7051030572507782, "eval_sickrelatedness_test_spearman": 0.7020350055407318, "eval_sickrelatedness_train_and_dev_spearman": 0.7082106452571797, "eval_sts12_spearman": 0.5997827794595306, "eval_sts13_spearman": 0.7980496603077974, "eval_sts14_spearman": 0.6948829895910905, "eval_sts15_spearman": 0.7853112805674828, "eval_sts16_spearman": 0.7659372500003789, "eval_stsbenchmark_spearman": 0.7502521515687675, "eval_stsbenchmark_test_spearman": 0.728894189985307, "eval_stsbenchmark_train_and_dev_spearman": 0.7544656214165828, "step": 11875 }, { "epoch": 0.8, "learning_rate": 7.802187826244837e-06, "loss": 0.0002, "step": 12500 }, { "epoch": 0.8, "eval_avg_sts": 0.7330359711356126, "eval_avg_test": 0.7291193278783638, "eval_avg_train_and_dev": 0.735211801074501, "eval_runtime": 20.8416, "eval_sickrelatedness_spearman": 0.7097679817259296, "eval_sickrelatedness_test_spearman": 0.7055788799611423, "eval_sickrelatedness_train_and_dev_spearman": 0.7139001197937711, "eval_sts12_spearman": 0.6106517927738067, "eval_sts13_spearman": 0.801722782519613, "eval_sts14_spearman": 0.697830283597753, "eval_sts15_spearman": 0.7905626238832419, "eval_sts16_spearman": 0.7686769836942824, "eval_stsbenchmark_spearman": 0.7520393497546612, "eval_stsbenchmark_test_spearman": 0.7288119487187074, "eval_stsbenchmark_train_and_dev_spearman": 0.7565234823552308, "step": 12500 }, { "epoch": 0.84, "learning_rate": 7.660342245018383e-06, "loss": 0.0001, "step": 13125 }, { "epoch": 0.84, "eval_avg_sts": 0.7378956830798117, "eval_avg_test": 0.733899347554639, "eval_avg_train_and_dev": 0.7387413140477561, "eval_runtime": 21.015, "eval_sickrelatedness_spearman": 0.710929018302664, "eval_sickrelatedness_test_spearman": 0.7063428540695924, "eval_sickrelatedness_train_and_dev_spearman": 0.7154190994394085, "eval_sts12_spearman": 0.6233547440533763, "eval_sts13_spearman": 0.8057484430966548, "eval_sts14_spearman": 0.7064661836743785, "eval_sts15_spearman": 0.788799289281052, "eval_sts16_spearman": 0.7725484698411716, "eval_stsbenchmark_spearman": 0.7574236333093849, "eval_stsbenchmark_test_spearman": 0.7340354488662467, "eval_stsbenchmark_train_and_dev_spearman": 0.7620635286561037, "step": 13125 }, { "epoch": 0.88, "learning_rate": 7.51849666379193e-06, "loss": 0.0001, "step": 13750 }, { "epoch": 0.88, "eval_avg_sts": 0.7378627077777866, "eval_avg_test": 0.7339054575865804, "eval_avg_train_and_dev": 0.7381744196853126, "eval_runtime": 20.944, "eval_sickrelatedness_spearman": 0.7089921810762922, "eval_sickrelatedness_test_spearman": 0.70509102387014, "eval_sickrelatedness_train_and_dev_spearman": 0.7128834324759727, "eval_sts12_spearman": 0.6228312321043947, "eval_sts13_spearman": 0.8058975575766927, "eval_sts14_spearman": 0.7070969399502143, "eval_sts15_spearman": 0.7893837860149883, "eval_sts16_spearman": 0.7720301633604001, "eval_stsbenchmark_spearman": 0.7588070943615248, "eval_stsbenchmark_test_spearman": 0.7350075002292331, "eval_stsbenchmark_train_and_dev_spearman": 0.7634654068946525, "step": 13750 }, { "epoch": 0.92, "learning_rate": 7.376651082565477e-06, "loss": 0.0001, "step": 14375 }, { "epoch": 0.92, "eval_avg_sts": 0.7359929925111067, "eval_avg_test": 0.7323951374834617, "eval_avg_train_and_dev": 0.7367857943216309, "eval_runtime": 20.9186, "eval_sickrelatedness_spearman": 0.7088303588545686, "eval_sickrelatedness_test_spearman": 0.7063407570964759, "eval_sickrelatedness_train_and_dev_spearman": 0.7113763758675955, "eval_sts12_spearman": 0.6136120965617352, "eval_sts13_spearman": 0.8054760021735932, "eval_sts14_spearman": 0.7046761197182577, "eval_sts15_spearman": 0.7905370837297951, "eval_sts16_spearman": 0.7711103256070158, "eval_stsbenchmark_spearman": 0.7577089609327817, "eval_stsbenchmark_test_spearman": 0.7350135774973598, "eval_stsbenchmark_train_and_dev_spearman": 0.7621952127756663, "step": 14375 }, { "epoch": 0.96, "learning_rate": 7.2348055013390225e-06, "loss": 0.0001, "step": 15000 }, { "epoch": 0.96, "eval_avg_sts": 0.7361943130200425, "eval_avg_test": 0.7321974439132747, "eval_avg_train_and_dev": 0.7382189598774902, "eval_runtime": 21.1353, "eval_sickrelatedness_spearman": 0.7096126852616816, "eval_sickrelatedness_test_spearman": 0.7056509931468986, "eval_sickrelatedness_train_and_dev_spearman": 0.7134760429404705, "eval_sts12_spearman": 0.613133445817732, "eval_sts13_spearman": 0.8048892508629485, "eval_sts14_spearman": 0.7046336569955289, "eval_sts15_spearman": 0.7904009982047565, "eval_sts16_spearman": 0.7725216175809143, "eval_stsbenchmark_spearman": 0.7581685364167358, "eval_stsbenchmark_test_spearman": 0.7341521447841448, "eval_stsbenchmark_train_and_dev_spearman": 0.7629618768145098, "step": 15000 }, { "epoch": 1.0, "learning_rate": 7.092959920112569e-06, "loss": 0.0001, "step": 15625 }, { "epoch": 1.0, "eval_avg_sts": 0.7375598442201905, "eval_avg_test": 0.7335423238398006, "eval_avg_train_and_dev": 0.7395439996671019, "eval_runtime": 20.8993, "eval_sickrelatedness_spearman": 0.707283647594327, "eval_sickrelatedness_test_spearman": 0.703616122718381, "eval_sickrelatedness_train_and_dev_spearman": 0.7109751567220617, "eval_sts12_spearman": 0.6159411441163902, "eval_sts13_spearman": 0.8049060590081953, "eval_sts14_spearman": 0.7108002254446067, "eval_sts15_spearman": 0.788937840589697, "eval_sts16_spearman": 0.7718528831999566, "eval_stsbenchmark_spearman": 0.7631971095881599, "eval_stsbenchmark_test_spearman": 0.7387419918013778, "eval_stsbenchmark_train_and_dev_spearman": 0.7681128426121422, "step": 15625 }, { "epoch": 1.04, "learning_rate": 6.951114338886115e-06, "loss": 0.0001, "step": 16250 }, { "epoch": 1.04, "eval_avg_sts": 0.739199635865537, "eval_avg_test": 0.7355897759739543, "eval_avg_train_and_dev": 0.739808186948654, "eval_runtime": 20.9838, "eval_sickrelatedness_spearman": 0.7080971683598302, "eval_sickrelatedness_test_spearman": 0.7051237540095253, "eval_sickrelatedness_train_and_dev_spearman": 0.7111032002738162, "eval_sts12_spearman": 0.6194643102288157, "eval_sts13_spearman": 0.8098252612446439, "eval_sts14_spearman": 0.7116464283311791, "eval_sts15_spearman": 0.7899553665444689, "eval_sts16_spearman": 0.7712889077310251, "eval_stsbenchmark_spearman": 0.7641200086187968, "eval_stsbenchmark_test_spearman": 0.7418244037280217, "eval_stsbenchmark_train_and_dev_spearman": 0.7685131736234918, "step": 16250 }, { "epoch": 1.08, "learning_rate": 6.809495710589624e-06, "loss": 0.0001, "step": 16875 }, { "epoch": 1.08, "eval_avg_sts": 0.7466792846411037, "eval_avg_test": 0.7432402664611477, "eval_avg_train_and_dev": 0.7444846866291048, "eval_runtime": 21.2164, "eval_sickrelatedness_spearman": 0.7071507310982136, "eval_sickrelatedness_test_spearman": 0.7036683201355792, "eval_sickrelatedness_train_and_dev_spearman": 0.7106803741065932, "eval_sts12_spearman": 0.6347617000535006, "eval_sts13_spearman": 0.8154858577332686, "eval_sts14_spearman": 0.7214692213354003, "eval_sts15_spearman": 0.7972123472428747, "eval_sts16_spearman": 0.7764992733505218, "eval_stsbenchmark_spearman": 0.7741758616739455, "eval_stsbenchmark_test_spearman": 0.7535851453768893, "eval_stsbenchmark_train_and_dev_spearman": 0.7782889991516164, "step": 16875 }, { "epoch": 1.12, "learning_rate": 6.66765012936317e-06, "loss": 0.0, "step": 17500 }, { "epoch": 1.12, "eval_avg_sts": 0.7445033109622727, "eval_avg_test": 0.7410841678179132, "eval_avg_train_and_dev": 0.7439110348985495, "eval_runtime": 20.7911, "eval_sickrelatedness_spearman": 0.7091854822762941, "eval_sickrelatedness_test_spearman": 0.7062734147342097, "eval_sickrelatedness_train_and_dev_spearman": 0.7121887413339406, "eval_sts12_spearman": 0.6279026674987728, "eval_sts13_spearman": 0.8113067283249641, "eval_sts14_spearman": 0.7213560371915237, "eval_sts15_spearman": 0.7972695297709387, "eval_sts16_spearman": 0.7730878646190888, "eval_stsbenchmark_spearman": 0.7714148670543273, "eval_stsbenchmark_test_spearman": 0.7503929325858947, "eval_stsbenchmark_train_and_dev_spearman": 0.7756333284631584, "step": 17500 }, { "epoch": 1.16, "learning_rate": 6.52603150106668e-06, "loss": 0.0001, "step": 18125 }, { "epoch": 1.16, "eval_avg_sts": 0.7424258185781643, "eval_avg_test": 0.7386605800972373, "eval_avg_train_and_dev": 0.7454667750890833, "eval_runtime": 21.0259, "eval_sickrelatedness_spearman": 0.7126108043398651, "eval_sickrelatedness_test_spearman": 0.7092073398325749, "eval_sickrelatedness_train_and_dev_spearman": 0.7160534091377377, "eval_sts12_spearman": 0.6254715777237362, "eval_sts13_spearman": 0.8088662475950285, "eval_sts14_spearman": 0.7138096529469777, "eval_sts15_spearman": 0.7925312951699506, "eval_sts16_spearman": 0.7732924074022455, "eval_stsbenchmark_spearman": 0.770398744869346, "eval_stsbenchmark_test_spearman": 0.7474455400101482, "eval_stsbenchmark_train_and_dev_spearman": 0.7748801410404289, "step": 18125 }, { "epoch": 1.2, "learning_rate": 6.384185919840225e-06, "loss": 0.0001, "step": 18750 }, { "epoch": 1.2, "eval_avg_sts": 0.7421874872103393, "eval_avg_test": 0.7386332506855652, "eval_avg_train_and_dev": 0.7448711406496367, "eval_runtime": 21.0153, "eval_sickrelatedness_spearman": 0.711420403952433, "eval_sickrelatedness_test_spearman": 0.7078637859078531, "eval_sickrelatedness_train_and_dev_spearman": 0.7149714012084747, "eval_sts12_spearman": 0.6198152391183793, "eval_sts13_spearman": 0.8112609489937583, "eval_sts14_spearman": 0.7137096793492043, "eval_sts15_spearman": 0.7909819007518449, "eval_sts16_spearman": 0.7775422386776029, "eval_stsbenchmark_spearman": 0.7705819996291529, "eval_stsbenchmark_test_spearman": 0.7492589620003135, "eval_stsbenchmark_train_and_dev_spearman": 0.7747708800907985, "step": 18750 }, { "epoch": 1.24, "learning_rate": 6.242340338613772e-06, "loss": 0.0001, "step": 19375 }, { "epoch": 1.24, "eval_avg_sts": 0.7433069618878057, "eval_avg_test": 0.7395454400266834, "eval_avg_train_and_dev": 0.7438536914759893, "eval_runtime": 20.8728, "eval_sickrelatedness_spearman": 0.7081170951655467, "eval_sickrelatedness_test_spearman": 0.7032372294807069, "eval_sickrelatedness_train_and_dev_spearman": 0.7129498291505492, "eval_sts12_spearman": 0.6319090861576243, "eval_sts13_spearman": 0.8105293738494971, "eval_sts14_spearman": 0.7133101665632953, "eval_sts15_spearman": 0.7957949511234824, "eval_sts16_spearman": 0.7729378020760672, "eval_stsbenchmark_spearman": 0.7705502582791273, "eval_stsbenchmark_test_spearman": 0.7490994709361111, "eval_stsbenchmark_train_and_dev_spearman": 0.7747575538014295, "step": 19375 }, { "epoch": 1.28, "learning_rate": 6.100494757387318e-06, "loss": 0.0001, "step": 20000 }, { "epoch": 1.28, "eval_avg_sts": 0.7499060859363368, "eval_avg_test": 0.7465428552421288, "eval_avg_train_and_dev": 0.7475216864388343, "eval_runtime": 20.8548, "eval_sickrelatedness_spearman": 0.7077244255701862, "eval_sickrelatedness_test_spearman": 0.7026718095093082, "eval_sickrelatedness_train_and_dev_spearman": 0.7128746659624327, "eval_sts12_spearman": 0.6428011959930444, "eval_sts13_spearman": 0.8168437631289807, "eval_sts14_spearman": 0.7226103243605893, "eval_sts15_spearman": 0.8024592797345551, "eval_sts16_spearman": 0.778377683273261, "eval_stsbenchmark_spearman": 0.778525929493741, "eval_stsbenchmark_test_spearman": 0.7600359306951637, "eval_stsbenchmark_train_and_dev_spearman": 0.7821687069152359, "step": 20000 }, { "epoch": 1.32, "learning_rate": 5.958649176160865e-06, "loss": 0.0, "step": 20625 }, { "epoch": 1.32, "eval_avg_sts": 0.7484118293338831, "eval_avg_test": 0.7449454204384204, "eval_avg_train_and_dev": 0.7464359816566652, "eval_runtime": 21.0711, "eval_sickrelatedness_spearman": 0.7074503178963312, "eval_sickrelatedness_test_spearman": 0.7023303881299484, "eval_sickrelatedness_train_and_dev_spearman": 0.7126512609959751, "eval_sts12_spearman": 0.6387436970256535, "eval_sts13_spearman": 0.8166418859715421, "eval_sts14_spearman": 0.7212103904818994, "eval_sts15_spearman": 0.8006237515601927, "eval_sts16_spearman": 0.7777887546493993, "eval_stsbenchmark_spearman": 0.7764240077521634, "eval_stsbenchmark_test_spearman": 0.7572790752503075, "eval_stsbenchmark_train_and_dev_spearman": 0.7802207023173555, "step": 20625 }, { "epoch": 1.36, "learning_rate": 5.816803594934411e-06, "loss": 0.0002, "step": 21250 }, { "epoch": 1.36, "eval_avg_sts": 0.7350019199078851, "eval_avg_test": 0.7318554174797746, "eval_avg_train_and_dev": 0.7352898672690771, "eval_runtime": 21.0366, "eval_sickrelatedness_spearman": 0.7074999541454676, "eval_sickrelatedness_test_spearman": 0.7046918538078386, "eval_sickrelatedness_train_and_dev_spearman": 0.7103029251438325, "eval_sts12_spearman": 0.6156570708780759, "eval_sts13_spearman": 0.8090436604472787, "eval_sts14_spearman": 0.7028681750157657, "eval_sts15_spearman": 0.7852035835160743, "eval_sts16_spearman": 0.7682223919418893, "eval_stsbenchmark_spearman": 0.7565186034106438, "eval_stsbenchmark_test_spearman": 0.737301186751499, "eval_stsbenchmark_train_and_dev_spearman": 0.7602768093943217, "step": 21250 }, { "epoch": 1.4, "learning_rate": 5.674958013707958e-06, "loss": 0.0001, "step": 21875 }, { "epoch": 1.4, "eval_avg_sts": 0.7380801801237084, "eval_avg_test": 0.7352880148005531, "eval_avg_train_and_dev": 0.7368474462537626, "eval_runtime": 20.8923, "eval_sickrelatedness_spearman": 0.7069077829822942, "eval_sickrelatedness_test_spearman": 0.7031570262809338, "eval_sickrelatedness_train_and_dev_spearman": 0.7106689319787851, "eval_sts12_spearman": 0.6231915518691485, "eval_sts13_spearman": 0.812850852612454, "eval_sts14_spearman": 0.7051573740637898, "eval_sts15_spearman": 0.7880917747314226, "eval_sts16_spearman": 0.7705225580148864, "eval_stsbenchmark_spearman": 0.7598393665919628, "eval_stsbenchmark_test_spearman": 0.7440449660312358, "eval_stsbenchmark_train_and_dev_spearman": 0.7630259605287402, "step": 21875 }, { "epoch": 1.44, "learning_rate": 5.533112432481503e-06, "loss": 0.0001, "step": 22500 }, { "epoch": 1.44, "eval_avg_sts": 0.7393794402273529, "eval_avg_test": 0.7363579691480349, "eval_avg_train_and_dev": 0.7395749061213661, "eval_runtime": 21.2448, "eval_sickrelatedness_spearman": 0.7102477930138079, "eval_sickrelatedness_test_spearman": 0.7064793937973988, "eval_sickrelatedness_train_and_dev_spearman": 0.7139973441863221, "eval_sts12_spearman": 0.6285417992352085, "eval_sts13_spearman": 0.8114621677478028, "eval_sts14_spearman": 0.7062564240963956, "eval_sts15_spearman": 0.7886571693369652, "eval_sts16_spearman": 0.7687956439175224, "eval_stsbenchmark_spearman": 0.7616950842437674, "eval_stsbenchmark_test_spearman": 0.7443131859049517, "eval_stsbenchmark_train_and_dev_spearman": 0.76515246805641, "step": 22500 }, { "epoch": 1.48, "learning_rate": 5.391266851255051e-06, "loss": 0.0001, "step": 23125 }, { "epoch": 1.48, "eval_avg_sts": 0.7388169479527533, "eval_avg_test": 0.7353151766024301, "eval_avg_train_and_dev": 0.7396876104946917, "eval_runtime": 20.7811, "eval_sickrelatedness_spearman": 0.7059146816791875, "eval_sickrelatedness_test_spearman": 0.7009398671844066, "eval_sickrelatedness_train_and_dev_spearman": 0.7109578919652116, "eval_sts12_spearman": 0.6337385819368979, "eval_sts13_spearman": 0.8077158631122653, "eval_sts14_spearman": 0.7041881431485422, "eval_sts15_spearman": 0.7897395308324451, "eval_sts16_spearman": 0.7659376148408716, "eval_stsbenchmark_spearman": 0.7644842201190634, "eval_stsbenchmark_test_spearman": 0.7449466351615813, "eval_stsbenchmark_train_and_dev_spearman": 0.7684173290241716, "step": 23125 }, { "epoch": 1.52, "learning_rate": 5.249648222958559e-06, "loss": 0.0001, "step": 23750 }, { "epoch": 1.52, "eval_avg_sts": 0.737637406431803, "eval_avg_test": 0.7340377292492831, "eval_avg_train_and_dev": 0.7393379830475855, "eval_runtime": 20.8525, "eval_sickrelatedness_spearman": 0.7079463598817589, "eval_sickrelatedness_test_spearman": 0.7029335485554905, "eval_sickrelatedness_train_and_dev_spearman": 0.7129341890695166, "eval_sts12_spearman": 0.6292541229208708, "eval_sts13_spearman": 0.806837478246872, "eval_sts14_spearman": 0.7027686303180887, "eval_sts15_spearman": 0.7886320383460999, "eval_sts16_spearman": 0.7662443034073775, "eval_stsbenchmark_spearman": 0.7617789119015519, "eval_stsbenchmark_test_spearman": 0.7415939829501823, "eval_stsbenchmark_train_and_dev_spearman": 0.7657417770256544, "step": 23750 }, { "epoch": 1.56, "learning_rate": 5.107802641732104e-06, "loss": 0.0001, "step": 24375 }, { "epoch": 1.56, "eval_avg_sts": 0.7400432752757417, "eval_avg_test": 0.7365741296789751, "eval_avg_train_and_dev": 0.7403337201879336, "eval_runtime": 21.0229, "eval_sickrelatedness_spearman": 0.706139895856204, "eval_sickrelatedness_test_spearman": 0.7011363476050757, "eval_sickrelatedness_train_and_dev_spearman": 0.7111677563573823, "eval_sts12_spearman": 0.6393107034863543, "eval_sts13_spearman": 0.8060305176934567, "eval_sts14_spearman": 0.707469456143251, "eval_sts15_spearman": 0.7884280597687867, "eval_sts16_spearman": 0.7672135350116814, "eval_stsbenchmark_spearman": 0.7657107589704576, "eval_stsbenchmark_test_spearman": 0.7464302880442203, "eval_stsbenchmark_train_and_dev_spearman": 0.7694996840184851, "step": 24375 }, { "epoch": 1.6, "learning_rate": 4.965957060505652e-06, "loss": 0.0001, "step": 25000 }, { "epoch": 1.6, "eval_avg_sts": 0.7420831068144936, "eval_avg_test": 0.7386039201251836, "eval_avg_train_and_dev": 0.7419238844699756, "eval_runtime": 20.7399, "eval_sickrelatedness_spearman": 0.7086802746453288, "eval_sickrelatedness_test_spearman": 0.703562769385137, "eval_sickrelatedness_train_and_dev_spearman": 0.7138105143408351, "eval_sts12_spearman": 0.643289234865721, "eval_sts13_spearman": 0.8077408521902738, "eval_sts14_spearman": 0.7094079968345621, "eval_sts15_spearman": 0.7910146281134862, "eval_sts16_spearman": 0.768300387542403, "eval_stsbenchmark_spearman": 0.7661483735096799, "eval_stsbenchmark_test_spearman": 0.7469115719447025, "eval_stsbenchmark_train_and_dev_spearman": 0.770037254599116, "step": 25000 }, { "epoch": 1.64, "learning_rate": 4.824111479279197e-06, "loss": 0.0001, "step": 25625 }, { "epoch": 1.64, "eval_avg_sts": 0.738553955422263, "eval_avg_test": 0.735270462011063, "eval_avg_train_and_dev": 0.74132969124333, "eval_runtime": 21.0644, "eval_sickrelatedness_spearman": 0.7107616379619491, "eval_sickrelatedness_test_spearman": 0.7068872624845219, "eval_sickrelatedness_train_and_dev_spearman": 0.714676339410284, "eval_sts12_spearman": 0.6304124246015519, "eval_sts13_spearman": 0.8044063778132966, "eval_sts14_spearman": 0.7046730335095999, "eval_sts15_spearman": 0.790737193065777, "eval_sts16_spearman": 0.764825533590085, "eval_stsbenchmark_spearman": 0.7640614874135817, "eval_stsbenchmark_test_spearman": 0.744951409012609, "eval_stsbenchmark_train_and_dev_spearman": 0.767983043076376, "step": 25625 }, { "epoch": 1.68, "learning_rate": 4.682492850982707e-06, "loss": 0.0001, "step": 26250 }, { "epoch": 1.68, "eval_avg_sts": 0.7337378478994732, "eval_avg_test": 0.7302481798812982, "eval_avg_train_and_dev": 0.7376796598903316, "eval_runtime": 20.8597, "eval_sickrelatedness_spearman": 0.7093236435437729, "eval_sickrelatedness_test_spearman": 0.7057306150582766, "eval_sickrelatedness_train_and_dev_spearman": 0.7130156557412384, "eval_sts12_spearman": 0.617871248735412, "eval_sts13_spearman": 0.8025593367032153, "eval_sts14_spearman": 0.7006950877629325, "eval_sts15_spearman": 0.785270135862098, "eval_sts16_spearman": 0.7622963644366769, "eval_stsbenchmark_spearman": 0.7581491182522041, "eval_stsbenchmark_test_spearman": 0.7373144706104764, "eval_stsbenchmark_train_and_dev_spearman": 0.7623436640394248, "step": 26250 }, { "epoch": 1.72, "learning_rate": 4.5406472697562535e-06, "loss": 0.0001, "step": 26875 }, { "epoch": 1.72, "eval_avg_sts": 0.7361414839625063, "eval_avg_test": 0.7327260474324794, "eval_avg_train_and_dev": 0.7408386426761701, "eval_runtime": 21.0874, "eval_sickrelatedness_spearman": 0.7130553203175825, "eval_sickrelatedness_test_spearman": 0.7085041459895807, "eval_sickrelatedness_train_and_dev_spearman": 0.7176214412488239, "eval_sts12_spearman": 0.6206145648203354, "eval_sts13_spearman": 0.8054403345042884, "eval_sts14_spearman": 0.7043767904617289, "eval_sts15_spearman": 0.785161403859601, "eval_sts16_spearman": 0.7642432303964917, "eval_stsbenchmark_spearman": 0.7600987433775165, "eval_stsbenchmark_test_spearman": 0.7407418619953297, "eval_stsbenchmark_train_and_dev_spearman": 0.7640558441035162, "step": 26875 }, { "epoch": 1.76, "learning_rate": 4.398801688529799e-06, "loss": 0.0001, "step": 27500 }, { "epoch": 1.76, "eval_avg_sts": 0.7385647317858953, "eval_avg_test": 0.7350557597229346, "eval_avg_train_and_dev": 0.7429292984249471, "eval_runtime": 20.7225, "eval_sickrelatedness_spearman": 0.7123203429942663, "eval_sickrelatedness_test_spearman": 0.7077860589944194, "eval_sickrelatedness_train_and_dev_spearman": 0.7168975415894818, "eval_sts12_spearman": 0.6250000311651265, "eval_sts13_spearman": 0.8098932511821607, "eval_sts14_spearman": 0.7051633776714191, "eval_sts15_spearman": 0.787197723837324, "eval_sts16_spearman": 0.7654891602041461, "eval_stsbenchmark_spearman": 0.7648892354468243, "eval_stsbenchmark_test_spearman": 0.7448607150059462, "eval_stsbenchmark_train_and_dev_spearman": 0.7689610552604125, "step": 27500 }, { "epoch": 1.8, "learning_rate": 4.2569561073033456e-06, "loss": 0.0001, "step": 28125 }, { "epoch": 1.8, "eval_avg_sts": 0.736228210862442, "eval_avg_test": 0.7323225373743962, "eval_avg_train_and_dev": 0.7422242610156418, "eval_runtime": 21.1674, "eval_sickrelatedness_spearman": 0.7117315495344994, "eval_sickrelatedness_test_spearman": 0.7062794242883943, "eval_sickrelatedness_train_and_dev_spearman": 0.7173189540357382, "eval_sts12_spearman": 0.6181306965992698, "eval_sts13_spearman": 0.8063871249869112, "eval_sts14_spearman": 0.7026158508695518, "eval_sts15_spearman": 0.7867321226315551, "eval_sts16_spearman": 0.7652562131980097, "eval_stsbenchmark_spearman": 0.7627439182172969, "eval_stsbenchmark_test_spearman": 0.7408563290470811, "eval_stsbenchmark_train_and_dev_spearman": 0.7671295679955453, "step": 28125 }, { "epoch": 1.84, "learning_rate": 4.115110526076892e-06, "loss": 0.0001, "step": 28750 }, { "epoch": 1.84, "eval_avg_sts": 0.7372002276835504, "eval_avg_test": 0.7334893393874011, "eval_avg_train_and_dev": 0.7419188871398015, "eval_runtime": 21.0224, "eval_sickrelatedness_spearman": 0.7115479327768666, "eval_sickrelatedness_test_spearman": 0.7061596518024446, "eval_sickrelatedness_train_and_dev_spearman": 0.7169625169788694, "eval_sts12_spearman": 0.6201372334499261, "eval_sts13_spearman": 0.8081427416220239, "eval_sts14_spearman": 0.7058112992260926, "eval_sts15_spearman": 0.7871505208738109, "eval_sts16_spearman": 0.7648927226152069, "eval_stsbenchmark_spearman": 0.7627191432209255, "eval_stsbenchmark_test_spearman": 0.7421312061223029, "eval_stsbenchmark_train_and_dev_spearman": 0.7668752573007337, "step": 28750 }, { "epoch": 1.88, "learning_rate": 3.9732649448504385e-06, "loss": 0.0002, "step": 29375 }, { "epoch": 1.88, "eval_avg_sts": 0.7391379223370436, "eval_avg_test": 0.7356189035069137, "eval_avg_train_and_dev": 0.7435335345244203, "eval_runtime": 20.9593, "eval_sickrelatedness_spearman": 0.7106110928678779, "eval_sickrelatedness_test_spearman": 0.7046551635771077, "eval_sickrelatedness_train_and_dev_spearman": 0.7166877770505714, "eval_sts12_spearman": 0.6242789115119458, "eval_sts13_spearman": 0.8073323769451142, "eval_sts14_spearman": 0.7082426452989751, "eval_sts15_spearman": 0.7878887744476603, "eval_sts16_spearman": 0.7689858922579222, "eval_stsbenchmark_spearman": 0.7666257630298101, "eval_stsbenchmark_test_spearman": 0.7479485605096713, "eval_stsbenchmark_train_and_dev_spearman": 0.7703792919982693, "step": 29375 }, { "epoch": 1.92, "learning_rate": 3.831419363623985e-06, "loss": 0.0001, "step": 30000 }, { "epoch": 1.92, "eval_avg_sts": 0.7379153295754598, "eval_avg_test": 0.7345327453977518, "eval_avg_train_and_dev": 0.7418020175283047, "eval_runtime": 20.9305, "eval_sickrelatedness_spearman": 0.7095180409050843, "eval_sickrelatedness_test_spearman": 0.7035887655223291, "eval_sickrelatedness_train_and_dev_spearman": 0.7155211362349155, "eval_sts12_spearman": 0.6242234464567169, "eval_sts13_spearman": 0.8060611779379976, "eval_sts14_spearman": 0.704822019666627, "eval_sts15_spearman": 0.7862063807761319, "eval_sts16_spearman": 0.770166017643468, "eval_stsbenchmark_spearman": 0.7644102236421939, "eval_stsbenchmark_test_spearman": 0.7466614097809916, "eval_stsbenchmark_train_and_dev_spearman": 0.7680828988216938, "step": 30000 }, { "epoch": 1.96, "learning_rate": 3.6898007353274935e-06, "loss": 0.0002, "step": 30625 }, { "epoch": 1.96, "eval_avg_sts": 0.7414417844482861, "eval_avg_test": 0.7377227000737309, "eval_avg_train_and_dev": 0.7426887168127401, "eval_runtime": 20.9682, "eval_sickrelatedness_spearman": 0.7083850749498312, "eval_sickrelatedness_test_spearman": 0.7020411335449008, "eval_sickrelatedness_train_and_dev_spearman": 0.7147828329921959, "eval_sts12_spearman": 0.6398748755196823, "eval_sts13_spearman": 0.8091407530457168, "eval_sts14_spearman": 0.7086523177233283, "eval_sts15_spearman": 0.7887396499382829, "eval_sts16_spearman": 0.7687341048891995, "eval_stsbenchmark_spearman": 0.7665657150719614, "eval_stsbenchmark_test_spearman": 0.7468760658550058, "eval_stsbenchmark_train_and_dev_spearman": 0.7705946006332842, "step": 30625 }, { "epoch": 2.0, "learning_rate": 3.54795515410104e-06, "loss": 0.0001, "step": 31250 }, { "epoch": 2.0, "eval_avg_sts": 0.740485017285834, "eval_avg_test": 0.7370491200682708, "eval_avg_train_and_dev": 0.7429999990678977, "eval_runtime": 20.7964, "eval_sickrelatedness_spearman": 0.7097765510038309, "eval_sickrelatedness_test_spearman": 0.7039997719017956, "eval_sickrelatedness_train_and_dev_spearman": 0.715634333306455, "eval_sts12_spearman": 0.6346257316776542, "eval_sts13_spearman": 0.8098227751570725, "eval_sts14_spearman": 0.7070680432539359, "eval_sts15_spearman": 0.7867814254070365, "eval_sts16_spearman": 0.7686156462120806, "eval_stsbenchmark_spearman": 0.766704948289228, "eval_stsbenchmark_test_spearman": 0.7484304468683206, "eval_stsbenchmark_train_and_dev_spearman": 0.7703656648293403, "step": 31250 }, { "epoch": 2.04, "learning_rate": 3.4063365258045485e-06, "loss": 0.0002, "step": 31875 }, { "epoch": 2.04, "eval_avg_sts": 0.7410438701692595, "eval_avg_test": 0.7373422938815762, "eval_avg_train_and_dev": 0.7446059296458287, "eval_runtime": 20.9546, "eval_sickrelatedness_spearman": 0.7105618614355229, "eval_sickrelatedness_test_spearman": 0.7049301366548643, "eval_sickrelatedness_train_and_dev_spearman": 0.7162726117635637, "eval_sts12_spearman": 0.6306599951032117, "eval_sts13_spearman": 0.807157187196233, "eval_sts14_spearman": 0.709921780686096, "eval_sts15_spearman": 0.7882494996283947, "eval_sts16_spearman": 0.7718392477227314, "eval_stsbenchmark_spearman": 0.7689175194126273, "eval_stsbenchmark_test_spearman": 0.7486382101795024, "eval_stsbenchmark_train_and_dev_spearman": 0.7729392475280936, "step": 31875 }, { "epoch": 2.08, "learning_rate": 3.264490944578095e-06, "loss": 0.0002, "step": 32500 }, { "epoch": 2.08, "eval_avg_sts": 0.7406899068085288, "eval_avg_test": 0.737240537915657, "eval_avg_train_and_dev": 0.7446535659927934, "eval_runtime": 21.252, "eval_sickrelatedness_spearman": 0.7129782951458211, "eval_sickrelatedness_test_spearman": 0.707353500824854, "eval_sickrelatedness_train_and_dev_spearman": 0.718621422755247, "eval_sts12_spearman": 0.631196341321798, "eval_sts13_spearman": 0.8084161729926506, "eval_sts14_spearman": 0.7092196209507754, "eval_sts15_spearman": 0.7846086513351899, "eval_sts16_spearman": 0.7714800805004121, "eval_stsbenchmark_spearman": 0.7669301854130546, "eval_stsbenchmark_test_spearman": 0.748409397483919, "eval_stsbenchmark_train_and_dev_spearman": 0.7706857092303399, "step": 32500 }, { "epoch": 2.12, "learning_rate": 3.1226453633516414e-06, "loss": 0.0001, "step": 33125 }, { "epoch": 2.12, "eval_avg_sts": 0.741908611558358, "eval_avg_test": 0.7385184801792626, "eval_avg_train_and_dev": 0.745921273906644, "eval_runtime": 20.8772, "eval_sickrelatedness_spearman": 0.712568145774737, "eval_sickrelatedness_test_spearman": 0.7071536357667555, "eval_sickrelatedness_train_and_dev_spearman": 0.7180482272764128, "eval_sts12_spearman": 0.6276979028419505, "eval_sts13_spearman": 0.8109754368105846, "eval_sts14_spearman": 0.7100445056205082, "eval_sts15_spearman": 0.7892064188117265, "eval_sts16_spearman": 0.7727292323220317, "eval_stsbenchmark_spearman": 0.7701386387269664, "eval_stsbenchmark_test_spearman": 0.751822229081281, "eval_stsbenchmark_train_and_dev_spearman": 0.7737943205368751, "step": 33125 }, { "epoch": 2.16, "learning_rate": 2.980799782125188e-06, "loss": 0.0002, "step": 33750 }, { "epoch": 2.16, "eval_avg_sts": 0.7387749148969024, "eval_avg_test": 0.7355275285778003, "eval_avg_train_and_dev": 0.7427705356417595, "eval_runtime": 21.9616, "eval_sickrelatedness_spearman": 0.7116362724362297, "eval_sickrelatedness_test_spearman": 0.7063792375154563, "eval_sickrelatedness_train_and_dev_spearman": 0.7170108350529064, "eval_sts12_spearman": 0.6210675920148566, "eval_sts13_spearman": 0.809613648614316, "eval_sts14_spearman": 0.7068348739830796, "eval_sts15_spearman": 0.786987945092306, "eval_sts16_spearman": 0.7702320756630628, "eval_stsbenchmark_spearman": 0.765051996474467, "eval_stsbenchmark_test_spearman": 0.7475773271615254, "eval_stsbenchmark_train_and_dev_spearman": 0.7685302362306128, "step": 33750 }, { "epoch": 2.2, "learning_rate": 2.8389542008987334e-06, "loss": 0.0001, "step": 34375 }, { "epoch": 2.2, "eval_avg_sts": 0.7372500677744379, "eval_avg_test": 0.7342484305550541, "eval_avg_train_and_dev": 0.7427227214188237, "eval_runtime": 21.98, "eval_sickrelatedness_spearman": 0.7124431266877929, "eval_sickrelatedness_test_spearman": 0.7076623305284827, "eval_sickrelatedness_train_and_dev_spearman": 0.7173681019291466, "eval_sts12_spearman": 0.6203661142159484, "eval_sts13_spearman": 0.806796081405754, "eval_sts14_spearman": 0.7047306825878202, "eval_sts15_spearman": 0.7819332414401158, "eval_sts16_spearman": 0.7696611847093084, "eval_stsbenchmark_spearman": 0.764820043374326, "eval_stsbenchmark_test_spearman": 0.7485893789979496, "eval_stsbenchmark_train_and_dev_spearman": 0.7680773409085008, "step": 34375 }, { "epoch": 2.24, "learning_rate": 2.69710861967228e-06, "loss": 0.0001, "step": 35000 }, { "epoch": 2.24, "eval_avg_sts": 0.7387507850343802, "eval_avg_test": 0.7355342211965702, "eval_avg_train_and_dev": 0.7445443682185053, "eval_runtime": 22.2171, "eval_sickrelatedness_spearman": 0.7125033872391564, "eval_sickrelatedness_test_spearman": 0.7073772850432812, "eval_sickrelatedness_train_and_dev_spearman": 0.7177230874857234, "eval_sts12_spearman": 0.6247069431706919, "eval_sts13_spearman": 0.8046176559037094, "eval_sts14_spearman": 0.7071478374873903, "eval_sts15_spearman": 0.7836492394816458, "eval_sts16_spearman": 0.7707143442262178, "eval_stsbenchmark_spearman": 0.7679160877318494, "eval_stsbenchmark_test_spearman": 0.7505262430630557, "eval_stsbenchmark_train_and_dev_spearman": 0.7713656489512873, "step": 35000 }, { "epoch": 2.28, "learning_rate": 2.5554899913757893e-06, "loss": 0.0001, "step": 35625 }, { "epoch": 2.28, "eval_avg_sts": 0.7377723938824481, "eval_avg_test": 0.7345737983826125, "eval_avg_train_and_dev": 0.7435797752746427, "eval_runtime": 22.0951, "eval_sickrelatedness_spearman": 0.7123552031305164, "eval_sickrelatedness_test_spearman": 0.7072368789839194, "eval_sickrelatedness_train_and_dev_spearman": 0.7175580611123827, "eval_sts12_spearman": 0.6227870968906012, "eval_sts13_spearman": 0.8046259866606313, "eval_sts14_spearman": 0.7045085522245373, "eval_sts15_spearman": 0.7838932428340308, "eval_sts16_spearman": 0.7701283228301814, "eval_stsbenchmark_spearman": 0.7661083526066392, "eval_stsbenchmark_test_spearman": 0.7488365082543857, "eval_stsbenchmark_train_and_dev_spearman": 0.7696014894369028, "step": 35625 }, { "epoch": 2.32, "learning_rate": 2.4136444101493353e-06, "loss": 0.0, "step": 36250 }, { "epoch": 2.32, "eval_avg_sts": 0.7386155577954275, "eval_avg_test": 0.7353702532577143, "eval_avg_train_and_dev": 0.7437646103034667, "eval_runtime": 20.9396, "eval_sickrelatedness_spearman": 0.7116453653128161, "eval_sickrelatedness_test_spearman": 0.7065116410259358, "eval_sickrelatedness_train_and_dev_spearman": 0.7168516996463055, "eval_sts12_spearman": 0.6248182960777477, "eval_sts13_spearman": 0.8056761080437632, "eval_sts14_spearman": 0.7057631810447995, "eval_sts15_spearman": 0.7847325098633744, "eval_sts16_spearman": 0.7705122919119067, "eval_stsbenchmark_spearman": 0.7671611523135844, "eval_stsbenchmark_test_spearman": 0.7495777448364731, "eval_stsbenchmark_train_and_dev_spearman": 0.7706775209606281, "step": 36250 }, { "epoch": 2.36, "learning_rate": 2.2717988289228813e-06, "loss": 0.0001, "step": 36875 }, { "epoch": 2.36, "eval_avg_sts": 0.740149603021757, "eval_avg_test": 0.7367482300057596, "eval_avg_train_and_dev": 0.7453363908257578, "eval_runtime": 20.671, "eval_sickrelatedness_spearman": 0.7126227371950188, "eval_sickrelatedness_test_spearman": 0.7077303136706309, "eval_sickrelatedness_train_and_dev_spearman": 0.7176363979763565, "eval_sts12_spearman": 0.6274167410561244, "eval_sts13_spearman": 0.8069745763455181, "eval_sts14_spearman": 0.7076654105229765, "eval_sts15_spearman": 0.7863406428310873, "eval_sts16_spearman": 0.7707586061652102, "eval_stsbenchmark_spearman": 0.7692685070363641, "eval_stsbenchmark_test_spearman": 0.7503513194487701, "eval_stsbenchmark_train_and_dev_spearman": 0.7730363836751589, "step": 36875 }, { "epoch": 2.4, "learning_rate": 2.129953247696428e-06, "loss": 0.0001, "step": 37500 }, { "epoch": 2.4, "eval_avg_sts": 0.7395131963250486, "eval_avg_test": 0.7361937524639587, "eval_avg_train_and_dev": 0.7453992073695682, "eval_runtime": 20.4199, "eval_sickrelatedness_spearman": 0.7127770844855997, "eval_sickrelatedness_test_spearman": 0.7078060121991877, "eval_sickrelatedness_train_and_dev_spearman": 0.717903112924754, "eval_sts12_spearman": 0.6274978324241501, "eval_sts13_spearman": 0.805585095341077, "eval_sts14_spearman": 0.7071221897597898, "eval_sts15_spearman": 0.7836729881285205, "eval_sts16_spearman": 0.7706690961987493, "eval_stsbenchmark_spearman": 0.7692680879374545, "eval_stsbenchmark_test_spearman": 0.7510030531962361, "eval_stsbenchmark_train_and_dev_spearman": 0.7728953018143825, "step": 37500 }, { "epoch": 2.44, "learning_rate": 1.9883346193999368e-06, "loss": 0.0002, "step": 38125 }, { "epoch": 2.44, "eval_avg_sts": 0.7387624644292584, "eval_avg_test": 0.7352227757479233, "eval_avg_train_and_dev": 0.7425307943142997, "eval_runtime": 20.7947, "eval_sickrelatedness_spearman": 0.7102028117293744, "eval_sickrelatedness_test_spearman": 0.7047476872624795, "eval_sickrelatedness_train_and_dev_spearman": 0.7157212183265218, "eval_sts12_spearman": 0.6320284509392846, "eval_sts13_spearman": 0.8033937882675368, "eval_sts14_spearman": 0.7054126300177038, "eval_sts15_spearman": 0.78530200196467, "eval_sts16_spearman": 0.7695255766427133, "eval_stsbenchmark_spearman": 0.7654719914435268, "eval_stsbenchmark_test_spearman": 0.7461492951410754, "eval_stsbenchmark_train_and_dev_spearman": 0.7693403703020776, "step": 38125 }, { "epoch": 2.48, "learning_rate": 1.846715991103445e-06, "loss": 0.0001, "step": 38750 }, { "epoch": 2.48, "eval_avg_sts": 0.7374253093552523, "eval_avg_test": 0.7339404463063186, "eval_avg_train_and_dev": 0.7421398368400338, "eval_runtime": 20.4703, "eval_sickrelatedness_spearman": 0.7107438562185104, "eval_sickrelatedness_test_spearman": 0.7052514908771994, "eval_sickrelatedness_train_and_dev_spearman": 0.7163253665937409, "eval_sts12_spearman": 0.6286712483831545, "eval_sts13_spearman": 0.8022301065346743, "eval_sts14_spearman": 0.7048882397124776, "eval_sts15_spearman": 0.7831405592800499, "eval_sts16_spearman": 0.7680903822018597, "eval_stsbenchmark_spearman": 0.7642127731560395, "eval_stsbenchmark_test_spearman": 0.7453110971548148, "eval_stsbenchmark_train_and_dev_spearman": 0.7679543070863266, "step": 38750 }, { "epoch": 2.52, "learning_rate": 1.7048704098769918e-06, "loss": 0.0, "step": 39375 }, { "epoch": 2.52, "eval_avg_sts": 0.7368107941649409, "eval_avg_test": 0.7334652203074301, "eval_avg_train_and_dev": 0.741976935359566, "eval_runtime": 20.8182, "eval_sickrelatedness_spearman": 0.7104882673504456, "eval_sickrelatedness_test_spearman": 0.7051023764808654, "eval_sickrelatedness_train_and_dev_spearman": 0.7159384852490909, "eval_sts12_spearman": 0.6268768757514214, "eval_sts13_spearman": 0.8029190740819949, "eval_sts14_spearman": 0.7022475030486164, "eval_sts15_spearman": 0.7826530446059501, "eval_sts16_spearman": 0.7681144502337414, "eval_stsbenchmark_spearman": 0.7643763440824164, "eval_stsbenchmark_test_spearman": 0.7463432179494219, "eval_stsbenchmark_train_and_dev_spearman": 0.7680153854700411, "step": 39375 }, { "epoch": 2.56, "learning_rate": 1.5630248286505382e-06, "loss": 0.0001, "step": 40000 }, { "epoch": 2.56, "eval_avg_sts": 0.7381003659828818, "eval_avg_test": 0.7347244014162845, "eval_avg_train_and_dev": 0.7429134210348216, "eval_runtime": 20.3825, "eval_sickrelatedness_spearman": 0.710316206946611, "eval_sickrelatedness_test_spearman": 0.7046947189644344, "eval_sickrelatedness_train_and_dev_spearman": 0.7160195342787384, "eval_sts12_spearman": 0.6302424800213983, "eval_sts13_spearman": 0.803247786425164, "eval_sts14_spearman": 0.7037417379983866, "eval_sts15_spearman": 0.7831307580280404, "eval_sts16_spearman": 0.7698214390074158, "eval_stsbenchmark_spearman": 0.7662021534531573, "eval_stsbenchmark_test_spearman": 0.7481918894691519, "eval_stsbenchmark_train_and_dev_spearman": 0.7698073077909047, "step": 40000 }, { "epoch": 2.6, "learning_rate": 1.4211792474240842e-06, "loss": 0.0, "step": 40625 }, { "epoch": 2.6, "eval_avg_sts": 0.7384929628200189, "eval_avg_test": 0.7351061112883038, "eval_avg_train_and_dev": 0.7432958438543483, "eval_runtime": 20.8515, "eval_sickrelatedness_spearman": 0.7108589590676125, "eval_sickrelatedness_test_spearman": 0.7053178143923748, "eval_sickrelatedness_train_and_dev_spearman": 0.7164885700394883, "eval_sts12_spearman": 0.6291888769664764, "eval_sts13_spearman": 0.8041919909911799, "eval_sts14_spearman": 0.7045984873525695, "eval_sts15_spearman": 0.7839088787514867, "eval_sts16_spearman": 0.7702165266681544, "eval_stsbenchmark_spearman": 0.766487019942652, "eval_stsbenchmark_test_spearman": 0.7483202038958853, "eval_stsbenchmark_train_and_dev_spearman": 0.7701031176692084, "step": 40625 }, { "epoch": 2.64, "learning_rate": 1.2793336661976307e-06, "loss": 0.0002, "step": 41250 }, { "epoch": 2.64, "eval_avg_sts": 0.7371561236957961, "eval_avg_test": 0.7339566580465452, "eval_avg_train_and_dev": 0.7422583980781617, "eval_runtime": 21.234, "eval_sickrelatedness_spearman": 0.7099911047770248, "eval_sickrelatedness_test_spearman": 0.7044999663247513, "eval_sickrelatedness_train_and_dev_spearman": 0.7155628833300955, "eval_sts12_spearman": 0.6254860905536815, "eval_sts13_spearman": 0.8017523519499081, "eval_sts14_spearman": 0.7028572928547238, "eval_sts15_spearman": 0.7845378195212266, "eval_sts16_spearman": 0.7699394553459527, "eval_stsbenchmark_spearman": 0.7655287508680542, "eval_stsbenchmark_test_spearman": 0.7486236297755721, "eval_stsbenchmark_train_and_dev_spearman": 0.7689539128262279, "step": 41250 }, { "epoch": 2.68, "learning_rate": 1.1374880849711772e-06, "loss": 0.0001, "step": 41875 }, { "epoch": 2.68, "eval_avg_sts": 0.7370631114339077, "eval_avg_test": 0.7338136544671776, "eval_avg_train_and_dev": 0.7423301757389922, "eval_runtime": 21.6943, "eval_sickrelatedness_spearman": 0.7102428073706631, "eval_sickrelatedness_test_spearman": 0.7048236058355623, "eval_sickrelatedness_train_and_dev_spearman": 0.7157821202465618, "eval_sts12_spearman": 0.6236458167934951, "eval_sts13_spearman": 0.8023700660147888, "eval_sts14_spearman": 0.703084563331531, "eval_sts15_spearman": 0.7846832363439064, "eval_sts16_spearman": 0.7700794957874387, "eval_stsbenchmark_spearman": 0.7653357943955308, "eval_stsbenchmark_test_spearman": 0.7480087971635208, "eval_stsbenchmark_train_and_dev_spearman": 0.7688782312314226, "step": 41875 }, { "epoch": 2.72, "learning_rate": 9.956425037447234e-07, "loss": 0.0001, "step": 42500 }, { "epoch": 2.72, "eval_avg_sts": 0.7382956278861713, "eval_avg_test": 0.734991312963749, "eval_avg_train_and_dev": 0.7428571308285099, "eval_runtime": 21.0897, "eval_sickrelatedness_spearman": 0.7102825098008535, "eval_sickrelatedness_test_spearman": 0.7048535850875951, "eval_sickrelatedness_train_and_dev_spearman": 0.7157899403058504, "eval_sts12_spearman": 0.6265558778138477, "eval_sts13_spearman": 0.8033181918156761, "eval_sts14_spearman": 0.7042972985487412, "eval_sts15_spearman": 0.7861678530544808, "eval_sts16_spearman": 0.7711254897645943, "eval_stsbenchmark_spearman": 0.7663221744050053, "eval_stsbenchmark_test_spearman": 0.7486208946613078, "eval_stsbenchmark_train_and_dev_spearman": 0.7699243213511694, "step": 42500 }, { "epoch": 2.76, "learning_rate": 8.540238754482321e-07, "loss": 0.0001, "step": 43125 }, { "epoch": 2.76, "eval_avg_sts": 0.7377715580722582, "eval_avg_test": 0.7345108563336945, "eval_avg_train_and_dev": 0.7427287762752324, "eval_runtime": 22.1235, "eval_sickrelatedness_spearman": 0.7106771021682546, "eval_sickrelatedness_test_spearman": 0.705266686228856, "eval_sickrelatedness_train_and_dev_spearman": 0.716181272162837, "eval_sts12_spearman": 0.6239496815809444, "eval_sts13_spearman": 0.8039159573225746, "eval_sts14_spearman": 0.7040700656997784, "eval_sts15_spearman": 0.7851629218072443, "eval_sts16_spearman": 0.7708450368779131, "eval_stsbenchmark_spearman": 0.7657801410490988, "eval_stsbenchmark_test_spearman": 0.7483656448185506, "eval_stsbenchmark_train_and_dev_spearman": 0.7692762803876276, "step": 43125 }, { "epoch": 2.8, "learning_rate": 7.121782942217785e-07, "loss": 0.0, "step": 43750 }, { "epoch": 2.8, "eval_avg_sts": 0.7376141103751033, "eval_avg_test": 0.7344015898023963, "eval_avg_train_and_dev": 0.742544402311891, "eval_runtime": 21.1096, "eval_sickrelatedness_spearman": 0.7104606827644014, "eval_sickrelatedness_test_spearman": 0.7050746381321955, "eval_sickrelatedness_train_and_dev_spearman": 0.7159137410961507, "eval_sts12_spearman": 0.6239536269253912, "eval_sts13_spearman": 0.8036508752187724, "eval_sts14_spearman": 0.7037219130584743, "eval_sts15_spearman": 0.7847939731114079, "eval_sts16_spearman": 0.7709951067177447, "eval_stsbenchmark_spearman": 0.7657225948295306, "eval_stsbenchmark_test_spearman": 0.748620995452788, "eval_stsbenchmark_train_and_dev_spearman": 0.7691750635276314, "step": 43750 }, { "epoch": 2.84, "learning_rate": 5.703327129953248e-07, "loss": 0.0001, "step": 44375 }, { "epoch": 2.84, "eval_avg_sts": 0.7375152940520443, "eval_avg_test": 0.7343389041124367, "eval_avg_train_and_dev": 0.7423452174385804, "eval_runtime": 22.0931, "eval_sickrelatedness_spearman": 0.7104542745458605, "eval_sickrelatedness_test_spearman": 0.705098130115833, "eval_sickrelatedness_train_and_dev_spearman": 0.7158819129939784, "eval_sts12_spearman": 0.6246511970226402, "eval_sts13_spearman": 0.8034092134157865, "eval_sts14_spearman": 0.7037963451422589, "eval_sts15_spearman": 0.7839004357252624, "eval_sts16_spearman": 0.7710148359050526, "eval_stsbenchmark_spearman": 0.7653807566074485, "eval_stsbenchmark_test_spearman": 0.7485021714602236, "eval_stsbenchmark_train_and_dev_spearman": 0.7688085218831824, "step": 44375 }, { "epoch": 2.88, "learning_rate": 4.2848713176887115e-07, "loss": 0.0, "step": 45000 }, { "epoch": 2.88, "eval_avg_sts": 0.7373122545485458, "eval_avg_test": 0.7341422495112051, "eval_avg_train_and_dev": 0.7421676000848536, "eval_runtime": 21.0333, "eval_sickrelatedness_spearman": 0.7104022241403064, "eval_sickrelatedness_test_spearman": 0.7050898274710157, "eval_sickrelatedness_train_and_dev_spearman": 0.7157835004453044, "eval_sts12_spearman": 0.6238646284703343, "eval_sts13_spearman": 0.8032636806389909, "eval_sts14_spearman": 0.7035512511045346, "eval_sts15_spearman": 0.7839647615882837, "eval_sts16_spearman": 0.7710192044322828, "eval_stsbenchmark_spearman": 0.7651200314650877, "eval_stsbenchmark_test_spearman": 0.7482423928729935, "eval_stsbenchmark_train_and_dev_spearman": 0.768551699724403, "step": 45000 }, { "epoch": 2.92, "learning_rate": 2.8664155054241755e-07, "loss": 0.0001, "step": 45625 }, { "epoch": 2.92, "eval_avg_sts": 0.7376021690547506, "eval_avg_test": 0.7344364718587105, "eval_avg_train_and_dev": 0.7420987124013947, "eval_runtime": 20.9052, "eval_sickrelatedness_spearman": 0.7100355064557035, "eval_sickrelatedness_test_spearman": 0.7047251628468761, "eval_sickrelatedness_train_and_dev_spearman": 0.7154114262576732, "eval_sts12_spearman": 0.6245606765009919, "eval_sts13_spearman": 0.8038598745242214, "eval_sts14_spearman": 0.7038508600251897, "eval_sts15_spearman": 0.7847361674768141, "eval_sts16_spearman": 0.770818571348578, "eval_stsbenchmark_spearman": 0.765353527051755, "eval_stsbenchmark_test_spearman": 0.7485039902883023, "eval_stsbenchmark_train_and_dev_spearman": 0.7687859985451162, "step": 45625 }, { "epoch": 2.96, "learning_rate": 1.447959693159639e-07, "loss": 0.0001, "step": 46250 }, { "epoch": 2.96, "eval_avg_sts": 0.737426533458678, "eval_avg_test": 0.7342698102173869, "eval_avg_train_and_dev": 0.7419274116141954, "eval_runtime": 20.2997, "eval_sickrelatedness_spearman": 0.7099669691358331, "eval_sickrelatedness_test_spearman": 0.7046739918719009, "eval_sickrelatedness_train_and_dev_spearman": 0.7153334695992425, "eval_sts12_spearman": 0.6238235463591946, "eval_sts13_spearman": 0.8041066380633766, "eval_sts14_spearman": 0.7036262868879434, "eval_sts15_spearman": 0.7846043548200616, "eval_sts16_spearman": 0.770752451815878, "eval_stsbenchmark_spearman": 0.7651054871284582, "eval_stsbenchmark_test_spearman": 0.7483014017033541, "eval_stsbenchmark_train_and_dev_spearman": 0.7685213536291483, "step": 46250 }, { "epoch": 3.0, "learning_rate": 2.9503880895102356e-09, "loss": 0.0001, "step": 46875 }, { "epoch": 3.0, "eval_avg_sts": 0.7374977429972559, "eval_avg_test": 0.7343576643139398, "eval_avg_train_and_dev": 0.7419344676757816, "eval_runtime": 20.8011, "eval_sickrelatedness_spearman": 0.7099435140291807, "eval_sickrelatedness_test_spearman": 0.7046738346759807, "eval_sickrelatedness_train_and_dev_spearman": 0.7152876184590635, "eval_sts12_spearman": 0.6240417964491191, "eval_sts13_spearman": 0.8041867824864875, "eval_sts14_spearman": 0.703700075594037, "eval_sts15_spearman": 0.7846174552488144, "eval_sts16_spearman": 0.7708088556462593, "eval_stsbenchmark_spearman": 0.7651857215268932, "eval_stsbenchmark_test_spearman": 0.7484748500968808, "eval_stsbenchmark_train_and_dev_spearman": 0.7685813168924998, "step": 46875 }, { "epoch": 3.0, "step": 46875, "total_flos": 0.0, "train_loss": 0.00036472132327159247, "train_runtime": 8039.9662, "train_samples_per_second": 373.136, "train_steps_per_second": 5.83 } ], "max_steps": 46875, "num_train_epochs": 3, "total_flos": 0.0, "trial_name": null, "trial_params": null }