|
{ |
|
"best_metric": 0.5866261398176292, |
|
"best_model_checkpoint": "result/unsup/simcse-indobert-content-FakeCLSTrain-precise", |
|
"epoch": 3.0, |
|
"global_step": 756, |
|
"is_hyper_param_search": false, |
|
"is_local_process_zero": true, |
|
"is_world_process_zero": true, |
|
"log_history": [ |
|
{ |
|
"epoch": 0.08, |
|
"eval_CR": 73.16, |
|
"eval_FakeCLSDev": 0.9856897538637664, |
|
"eval_FakeCLSTest": 0.5, |
|
"eval_FakeCLSTrain": 0.5635838150289018, |
|
"eval_FakePairs": -0.0605484557984162, |
|
"eval_FakePairsNLI": 62.09, |
|
"eval_MPQA": 77.41, |
|
"eval_MR": 61.65, |
|
"eval_MRPC": 69.68, |
|
"eval_SST2": 69.95, |
|
"eval_SUBJ": 82.09, |
|
"eval_TREC": 65.04, |
|
"eval_avg_sts": 0.594301042267096, |
|
"eval_avg_transfer": 71.28285714285714, |
|
"eval_sickr_spearman": 0.5739718062787506, |
|
"eval_stsb_indo": 0.3209488720594688, |
|
"eval_stsb_spearman": 0.6146302782554416, |
|
"step": 20 |
|
}, |
|
{ |
|
"epoch": 0.16, |
|
"eval_CR": 72.46, |
|
"eval_FakeCLSDev": 0.9851598173515982, |
|
"eval_FakeCLSTest": 0.44155844155844154, |
|
"eval_FakeCLSTrain": 0.5580736543909348, |
|
"eval_FakePairs": -0.06458236398195215, |
|
"eval_FakePairsNLI": 62.56, |
|
"eval_MPQA": 77.78, |
|
"eval_MR": 61.94, |
|
"eval_MRPC": 69.92, |
|
"eval_SST2": 69.72, |
|
"eval_SUBJ": 81.95, |
|
"eval_TREC": 63.56, |
|
"eval_avg_sts": 0.5942119231714487, |
|
"eval_avg_transfer": 71.04714285714286, |
|
"eval_sickr_spearman": 0.572479335863764, |
|
"eval_stsb_indo": 0.33555082512700035, |
|
"eval_stsb_spearman": 0.6159445104791332, |
|
"step": 40 |
|
}, |
|
{ |
|
"epoch": 0.24, |
|
"eval_CR": 72.2, |
|
"eval_FakeCLSDev": 0.9883973894126178, |
|
"eval_FakeCLSTest": 0.49557522123893805, |
|
"eval_FakeCLSTrain": 0.5488126649076517, |
|
"eval_FakePairs": -0.06162683895274021, |
|
"eval_FakePairsNLI": 62.25, |
|
"eval_MPQA": 77.85, |
|
"eval_MR": 61.31, |
|
"eval_MRPC": 70.17, |
|
"eval_SST2": 69.95, |
|
"eval_SUBJ": 82.79, |
|
"eval_TREC": 63.96, |
|
"eval_avg_sts": 0.5899751766827189, |
|
"eval_avg_transfer": 71.17571428571428, |
|
"eval_sickr_spearman": 0.5702007884426414, |
|
"eval_stsb_indo": 0.31928216322520336, |
|
"eval_stsb_spearman": 0.6097495649227964, |
|
"step": 60 |
|
}, |
|
{ |
|
"epoch": 0.32, |
|
"eval_CR": 72.32, |
|
"eval_FakeCLSDev": 0.9828571428571429, |
|
"eval_FakeCLSTest": 0.4358974358974359, |
|
"eval_FakeCLSTrain": 0.505720823798627, |
|
"eval_FakePairs": -0.057193528168424795, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.43, |
|
"eval_MR": 62.04, |
|
"eval_MRPC": 69.38, |
|
"eval_SST2": 69.72, |
|
"eval_SUBJ": 82.67, |
|
"eval_TREC": 63.44, |
|
"eval_avg_sts": 0.6008571910373384, |
|
"eval_avg_transfer": 70.99999999999999, |
|
"eval_sickr_spearman": 0.5721567109554042, |
|
"eval_stsb_indo": 0.3125770393791657, |
|
"eval_stsb_spearman": 0.6295576711192724, |
|
"step": 80 |
|
}, |
|
{ |
|
"epoch": 0.4, |
|
"eval_CR": 71.95, |
|
"eval_FakeCLSDev": 1.0, |
|
"eval_FakeCLSTest": 0.4166666666666667, |
|
"eval_FakeCLSTrain": 0.5508474576271186, |
|
"eval_FakePairs": -0.05351908362129137, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.06, |
|
"eval_MR": 62.45, |
|
"eval_MRPC": 70.22, |
|
"eval_SST2": 69.04, |
|
"eval_SUBJ": 82.31, |
|
"eval_TREC": 63.44, |
|
"eval_avg_sts": 0.5941874894842695, |
|
"eval_avg_transfer": 70.92428571428572, |
|
"eval_sickr_spearman": 0.5704407998565277, |
|
"eval_stsb_indo": 0.30895400373418236, |
|
"eval_stsb_spearman": 0.6179341791120112, |
|
"step": 100 |
|
}, |
|
{ |
|
"epoch": 0.48, |
|
"eval_CR": 72.01, |
|
"eval_FakeCLSDev": 1.0, |
|
"eval_FakeCLSTest": 0.47058823529411764, |
|
"eval_FakeCLSTrain": 0.5430809399477807, |
|
"eval_FakePairs": -0.030993140962777686, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 76.94, |
|
"eval_MR": 61.46, |
|
"eval_MRPC": 69.68, |
|
"eval_SST2": 67.32, |
|
"eval_SUBJ": 81.42, |
|
"eval_TREC": 64.38, |
|
"eval_avg_sts": 0.5907439625277868, |
|
"eval_avg_transfer": 70.45857142857143, |
|
"eval_sickr_spearman": 0.5674561472131634, |
|
"eval_stsb_indo": 0.2897446195810389, |
|
"eval_stsb_spearman": 0.6140317778424101, |
|
"step": 120 |
|
}, |
|
{ |
|
"epoch": 0.56, |
|
"eval_CR": 72.67, |
|
"eval_FakeCLSDev": 0.9917948717948718, |
|
"eval_FakeCLSTest": 0.5957446808510638, |
|
"eval_FakeCLSTrain": 0.5555555555555556, |
|
"eval_FakePairs": -0.04577079838059695, |
|
"eval_FakePairsNLI": 63.18, |
|
"eval_MPQA": 77.25, |
|
"eval_MR": 61.59, |
|
"eval_MRPC": 69.87, |
|
"eval_SST2": 69.95, |
|
"eval_SUBJ": 81.79, |
|
"eval_TREC": 62.4, |
|
"eval_avg_sts": 0.601502215900739, |
|
"eval_avg_transfer": 70.78857142857143, |
|
"eval_sickr_spearman": 0.5684590846422926, |
|
"eval_stsb_indo": 0.3097638766411074, |
|
"eval_stsb_spearman": 0.6345453471591856, |
|
"step": 140 |
|
}, |
|
{ |
|
"epoch": 0.63, |
|
"eval_CR": 72.47, |
|
"eval_FakeCLSDev": 0.9908735332464146, |
|
"eval_FakeCLSTest": 0.5632911392405063, |
|
"eval_FakeCLSTrain": 0.556786703601108, |
|
"eval_FakePairs": -0.05599533972914216, |
|
"eval_FakePairsNLI": 60.69, |
|
"eval_MPQA": 77.78, |
|
"eval_MR": 61.95, |
|
"eval_MRPC": 69.87, |
|
"eval_SST2": 70.3, |
|
"eval_SUBJ": 81.82, |
|
"eval_TREC": 61.57, |
|
"eval_avg_sts": 0.5932109567201365, |
|
"eval_avg_transfer": 70.82285714285715, |
|
"eval_sickr_spearman": 0.5648520449864932, |
|
"eval_stsb_indo": 0.31916108185452513, |
|
"eval_stsb_spearman": 0.6215698684537796, |
|
"step": 160 |
|
}, |
|
{ |
|
"epoch": 0.71, |
|
"eval_CR": 72.56, |
|
"eval_FakeCLSDev": 0.9912587412587412, |
|
"eval_FakeCLSTest": 0.5294117647058824, |
|
"eval_FakeCLSTrain": 0.5282555282555282, |
|
"eval_FakePairs": -0.06366374574055107, |
|
"eval_FakePairsNLI": 63.96, |
|
"eval_MPQA": 77.65, |
|
"eval_MR": 62.88, |
|
"eval_MRPC": 69.85, |
|
"eval_SST2": 68.69, |
|
"eval_SUBJ": 82.11, |
|
"eval_TREC": 64.78, |
|
"eval_avg_sts": 0.5965775256163806, |
|
"eval_avg_transfer": 71.21714285714287, |
|
"eval_sickr_spearman": 0.5614600886029337, |
|
"eval_stsb_indo": 0.331965762222026, |
|
"eval_stsb_spearman": 0.6316949626298273, |
|
"step": 180 |
|
}, |
|
{ |
|
"epoch": 0.79, |
|
"eval_CR": 72.59, |
|
"eval_FakeCLSDev": 0.9902200488997555, |
|
"eval_FakeCLSTest": 0.5392670157068062, |
|
"eval_FakeCLSTrain": 0.555858310626703, |
|
"eval_FakePairs": -0.06334422882340904, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.32, |
|
"eval_MR": 62.44, |
|
"eval_MRPC": 70.41, |
|
"eval_SST2": 69.72, |
|
"eval_SUBJ": 82.33, |
|
"eval_TREC": 62.16, |
|
"eval_avg_sts": 0.5994183846840018, |
|
"eval_avg_transfer": 70.99571428571427, |
|
"eval_sickr_spearman": 0.5617657105013845, |
|
"eval_stsb_indo": 0.3362704444969455, |
|
"eval_stsb_spearman": 0.6370710588666191, |
|
"step": 200 |
|
}, |
|
{ |
|
"epoch": 0.87, |
|
"eval_CR": 71.93, |
|
"eval_FakeCLSDev": 0.9864066193853428, |
|
"eval_FakeCLSTest": 0.44529750479846447, |
|
"eval_FakeCLSTrain": 0.4653846153846154, |
|
"eval_FakePairs": -0.0781218862412283, |
|
"eval_FakePairsNLI": 60.69, |
|
"eval_MPQA": 77.74, |
|
"eval_MR": 62.75, |
|
"eval_MRPC": 69.95, |
|
"eval_SST2": 69.72, |
|
"eval_SUBJ": 81.79, |
|
"eval_TREC": 62.12, |
|
"eval_avg_sts": 0.6017830377412869, |
|
"eval_avg_transfer": 70.85714285714286, |
|
"eval_sickr_spearman": 0.5577540568470704, |
|
"eval_stsb_indo": 0.3330327185293657, |
|
"eval_stsb_spearman": 0.6458120186355034, |
|
"step": 220 |
|
}, |
|
{ |
|
"epoch": 0.95, |
|
"eval_CR": 72.07, |
|
"eval_FakeCLSDev": 0.9851767388825542, |
|
"eval_FakeCLSTest": 0.4206081081081081, |
|
"eval_FakeCLSTrain": 0.5619596541786743, |
|
"eval_FakePairs": -0.06254543653055394, |
|
"eval_FakePairsNLI": 62.4, |
|
"eval_MPQA": 77.63, |
|
"eval_MR": 62.42, |
|
"eval_MRPC": 70.39, |
|
"eval_SST2": 67.89, |
|
"eval_SUBJ": 81.1, |
|
"eval_TREC": 63.5, |
|
"eval_avg_sts": 0.6093138534071756, |
|
"eval_avg_transfer": 70.71428571428571, |
|
"eval_sickr_spearman": 0.5638283100904413, |
|
"eval_stsb_indo": 0.3179406446290444, |
|
"eval_stsb_spearman": 0.6547993967239099, |
|
"step": 240 |
|
}, |
|
{ |
|
"epoch": 1.03, |
|
"eval_CR": 71.91, |
|
"eval_FakeCLSDev": 0.9855156431054461, |
|
"eval_FakeCLSTest": 0.4250871080139373, |
|
"eval_FakeCLSTrain": 0.5616045845272206, |
|
"eval_FakePairs": -0.0714120309812455, |
|
"eval_FakePairsNLI": 62.71, |
|
"eval_MPQA": 77.76, |
|
"eval_MR": 62.24, |
|
"eval_MRPC": 70.29, |
|
"eval_SST2": 69.27, |
|
"eval_SUBJ": 81.23, |
|
"eval_TREC": 62.22, |
|
"eval_avg_sts": 0.6098143668968183, |
|
"eval_avg_transfer": 70.70285714285714, |
|
"eval_sickr_spearman": 0.5708000244641833, |
|
"eval_stsb_indo": 0.3233259049654498, |
|
"eval_stsb_spearman": 0.6488287093294535, |
|
"step": 260 |
|
}, |
|
{ |
|
"epoch": 1.11, |
|
"eval_CR": 71.6, |
|
"eval_FakeCLSDev": 0.9872, |
|
"eval_FakeCLSTest": 0.4154929577464789, |
|
"eval_FakeCLSTrain": 0.5055187637969095, |
|
"eval_FakePairs": -0.06158689255421782, |
|
"eval_FakePairsNLI": 61.31, |
|
"eval_MPQA": 77.74, |
|
"eval_MR": 62.29, |
|
"eval_MRPC": 69.85, |
|
"eval_SST2": 69.84, |
|
"eval_SUBJ": 81.13, |
|
"eval_TREC": 61.83, |
|
"eval_avg_sts": 0.6152679773595253, |
|
"eval_avg_transfer": 70.61142857142856, |
|
"eval_sickr_spearman": 0.5785564709732213, |
|
"eval_stsb_indo": 0.3262186264651655, |
|
"eval_stsb_spearman": 0.6519794837458294, |
|
"step": 280 |
|
}, |
|
{ |
|
"epoch": 1.19, |
|
"eval_CR": 71.99, |
|
"eval_FakeCLSDev": 0.9887798036465638, |
|
"eval_FakeCLSTest": 0.3881118881118881, |
|
"eval_FakeCLSTrain": 0.5628415300546448, |
|
"eval_FakePairs": -0.05176174057701016, |
|
"eval_FakePairsNLI": 63.18, |
|
"eval_MPQA": 77.82, |
|
"eval_MR": 62.44, |
|
"eval_MRPC": 70.49, |
|
"eval_SST2": 69.38, |
|
"eval_SUBJ": 81.6, |
|
"eval_TREC": 62.64, |
|
"eval_avg_sts": 0.6081735394724535, |
|
"eval_avg_transfer": 70.90857142857143, |
|
"eval_sickr_spearman": 0.5737140233573322, |
|
"eval_stsb_indo": 0.33079265734267466, |
|
"eval_stsb_spearman": 0.6426330555875746, |
|
"step": 300 |
|
}, |
|
{ |
|
"epoch": 1.27, |
|
"eval_CR": 71.69, |
|
"eval_FakeCLSDev": 0.9950738916256158, |
|
"eval_FakeCLSTest": 0.32575757575757575, |
|
"eval_FakeCLSTrain": 0.5660919540229885, |
|
"eval_FakePairs": -0.03634504932490683, |
|
"eval_FakePairsNLI": 61.31, |
|
"eval_MPQA": 78.3, |
|
"eval_MR": 61.9, |
|
"eval_MRPC": 70.41, |
|
"eval_SST2": 68.81, |
|
"eval_SUBJ": 81.27, |
|
"eval_TREC": 64.03, |
|
"eval_avg_sts": 0.6034605442857701, |
|
"eval_avg_transfer": 70.91571428571429, |
|
"eval_sickr_spearman": 0.5857587346338663, |
|
"eval_stsb_indo": 0.3175442989614004, |
|
"eval_stsb_spearman": 0.621162353937674, |
|
"step": 320 |
|
}, |
|
{ |
|
"epoch": 1.35, |
|
"eval_CR": 72.16, |
|
"eval_FakeCLSDev": 0.8333333333333334, |
|
"eval_FakeCLSTest": 0.2903225806451613, |
|
"eval_FakeCLSTrain": 0.5706051873198847, |
|
"eval_FakePairs": -0.009345869826404611, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.78, |
|
"eval_MR": 61.92, |
|
"eval_MRPC": 69.87, |
|
"eval_SST2": 68.35, |
|
"eval_SUBJ": 81.05, |
|
"eval_TREC": 63.74, |
|
"eval_avg_sts": 0.5978999565018066, |
|
"eval_avg_transfer": 70.69571428571429, |
|
"eval_sickr_spearman": 0.5846008969025991, |
|
"eval_stsb_indo": 0.2896616256780842, |
|
"eval_stsb_spearman": 0.6111990161010141, |
|
"step": 340 |
|
}, |
|
{ |
|
"epoch": 1.43, |
|
"eval_CR": 72.4, |
|
"eval_FakeCLSDev": 0.75, |
|
"eval_FakeCLSTest": 0.6, |
|
"eval_FakeCLSTrain": 0.5866261398176292, |
|
"eval_FakePairs": 0.009186111367833594, |
|
"eval_FakePairsNLI": 61.31, |
|
"eval_MPQA": 77.71, |
|
"eval_MR": 62.08, |
|
"eval_MRPC": 70.66, |
|
"eval_SST2": 68.92, |
|
"eval_SUBJ": 81.17, |
|
"eval_TREC": 62.38, |
|
"eval_avg_sts": 0.6025418380141683, |
|
"eval_avg_transfer": 70.76, |
|
"eval_sickr_spearman": 0.5790834682182003, |
|
"eval_stsb_indo": 0.2856567184812426, |
|
"eval_stsb_spearman": 0.6260002078101363, |
|
"step": 360 |
|
}, |
|
{ |
|
"epoch": 1.51, |
|
"eval_CR": 71.56, |
|
"eval_FakeCLSDev": 1.0, |
|
"eval_FakeCLSTest": 0.4, |
|
"eval_FakeCLSTrain": 0.5331695331695332, |
|
"eval_FakePairs": 0.01469777818853375, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.58, |
|
"eval_MR": 62.1, |
|
"eval_MRPC": 69.77, |
|
"eval_SST2": 69.5, |
|
"eval_SUBJ": 80.9, |
|
"eval_TREC": 60.66, |
|
"eval_avg_sts": 0.6008793834242059, |
|
"eval_avg_transfer": 70.29571428571428, |
|
"eval_sickr_spearman": 0.5750952057102866, |
|
"eval_stsb_indo": 0.2851042737082779, |
|
"eval_stsb_spearman": 0.626663561138125, |
|
"step": 380 |
|
}, |
|
{ |
|
"epoch": 1.59, |
|
"eval_CR": 72.58, |
|
"eval_FakeCLSDev": 0.75, |
|
"eval_FakeCLSTest": 0.2558139534883721, |
|
"eval_FakeCLSTrain": 0.5617977528089888, |
|
"eval_FakePairs": -0.02212654651208613, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.56, |
|
"eval_MR": 61.98, |
|
"eval_MRPC": 70.56, |
|
"eval_SST2": 69.5, |
|
"eval_SUBJ": 81.99, |
|
"eval_TREC": 60.35, |
|
"eval_avg_sts": 0.6094671498051143, |
|
"eval_avg_transfer": 70.64571428571429, |
|
"eval_sickr_spearman": 0.5732523003792077, |
|
"eval_stsb_indo": 0.29926655176105116, |
|
"eval_stsb_spearman": 0.645681999231021, |
|
"step": 400 |
|
}, |
|
{ |
|
"epoch": 1.67, |
|
"eval_CR": 72.26, |
|
"eval_FakeCLSDev": 0.75, |
|
"eval_FakeCLSTest": 0.24489795918367346, |
|
"eval_FakeCLSTrain": 0.5461538461538461, |
|
"eval_FakePairs": -0.04640983221488102, |
|
"eval_FakePairsNLI": 62.56, |
|
"eval_MPQA": 78.1, |
|
"eval_MR": 62.29, |
|
"eval_MRPC": 69.97, |
|
"eval_SST2": 70.07, |
|
"eval_SUBJ": 82.35, |
|
"eval_TREC": 58.6, |
|
"eval_avg_sts": 0.6147454616026657, |
|
"eval_avg_transfer": 70.52, |
|
"eval_sickr_spearman": 0.5760541946815999, |
|
"eval_stsb_indo": 0.3099846776429508, |
|
"eval_stsb_spearman": 0.6534367285237316, |
|
"step": 420 |
|
}, |
|
{ |
|
"epoch": 1.75, |
|
"eval_CR": 72.46, |
|
"eval_FakeCLSDev": 0.6666666666666666, |
|
"eval_FakeCLSTest": 0.2564102564102564, |
|
"eval_FakeCLSTrain": 0.5638888888888889, |
|
"eval_FakePairs": -0.05903075044199152, |
|
"eval_FakePairsNLI": 60.84, |
|
"eval_MPQA": 77.68, |
|
"eval_MR": 62.84, |
|
"eval_MRPC": 70.51, |
|
"eval_SST2": 70.18, |
|
"eval_SUBJ": 82.59, |
|
"eval_TREC": 60.55, |
|
"eval_avg_sts": 0.6163739495549794, |
|
"eval_avg_transfer": 70.97285714285715, |
|
"eval_sickr_spearman": 0.5752962639009065, |
|
"eval_stsb_indo": 0.3144443089302872, |
|
"eval_stsb_spearman": 0.6574516352090524, |
|
"step": 440 |
|
}, |
|
{ |
|
"epoch": 1.83, |
|
"eval_CR": 72.66, |
|
"eval_FakeCLSDev": 0.8, |
|
"eval_FakeCLSTest": 0.21621621621621623, |
|
"eval_FakeCLSTrain": 0.5635838150289018, |
|
"eval_FakePairs": -0.06614000184840188, |
|
"eval_FakePairsNLI": 61.62, |
|
"eval_MPQA": 77.6, |
|
"eval_MR": 63.18, |
|
"eval_MRPC": 70.58, |
|
"eval_SST2": 68.69, |
|
"eval_SUBJ": 82.96, |
|
"eval_TREC": 60.49, |
|
"eval_avg_sts": 0.6191142419028671, |
|
"eval_avg_transfer": 70.88, |
|
"eval_sickr_spearman": 0.5765413741434867, |
|
"eval_stsb_indo": 0.32117950245618027, |
|
"eval_stsb_spearman": 0.6616871096622475, |
|
"step": 460 |
|
}, |
|
{ |
|
"epoch": 1.9, |
|
"eval_CR": 72.22, |
|
"eval_FakeCLSDev": 0.9850746268656716, |
|
"eval_FakeCLSTest": 0.2713178294573643, |
|
"eval_FakeCLSTrain": 0.5596590909090909, |
|
"eval_FakePairs": -0.06198628192555537, |
|
"eval_FakePairsNLI": 62.4, |
|
"eval_MPQA": 77.47, |
|
"eval_MR": 63.01, |
|
"eval_MRPC": 70.22, |
|
"eval_SST2": 69.04, |
|
"eval_SUBJ": 82.91, |
|
"eval_TREC": 60.24, |
|
"eval_avg_sts": 0.6214756182164681, |
|
"eval_avg_transfer": 70.73, |
|
"eval_sickr_spearman": 0.5775538697618021, |
|
"eval_stsb_indo": 0.32220638413969804, |
|
"eval_stsb_spearman": 0.6653973666711341, |
|
"step": 480 |
|
}, |
|
{ |
|
"epoch": 1.98, |
|
"learning_rate": 1.693121693121693e-05, |
|
"loss": 0.039, |
|
"step": 500 |
|
}, |
|
{ |
|
"epoch": 1.98, |
|
"eval_CR": 72.6, |
|
"eval_FakeCLSDev": 0.9814814814814815, |
|
"eval_FakeCLSTest": 0.23529411764705882, |
|
"eval_FakeCLSTrain": 0.5516304347826086, |
|
"eval_FakePairs": -0.05535630589485809, |
|
"eval_FakePairsNLI": 61.31, |
|
"eval_MPQA": 77.9, |
|
"eval_MR": 62.95, |
|
"eval_MRPC": 70.24, |
|
"eval_SST2": 69.27, |
|
"eval_SUBJ": 82.79, |
|
"eval_TREC": 61.37, |
|
"eval_avg_sts": 0.623177260213142, |
|
"eval_avg_transfer": 71.01714285714286, |
|
"eval_sickr_spearman": 0.5778529113993559, |
|
"eval_stsb_indo": 0.32376765191924134, |
|
"eval_stsb_spearman": 0.668501609026928, |
|
"step": 500 |
|
}, |
|
{ |
|
"epoch": 2.06, |
|
"eval_CR": 72.46, |
|
"eval_FakeCLSDev": 0.9794871794871794, |
|
"eval_FakeCLSTest": 0.29891304347826086, |
|
"eval_FakeCLSTrain": 0.5808383233532934, |
|
"eval_FakePairs": -0.046010436068453474, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.7, |
|
"eval_MR": 63.16, |
|
"eval_MRPC": 70.07, |
|
"eval_SST2": 69.72, |
|
"eval_SUBJ": 82.93, |
|
"eval_TREC": 61.23, |
|
"eval_avg_sts": 0.6209015004562825, |
|
"eval_avg_transfer": 71.03857142857143, |
|
"eval_sickr_spearman": 0.5749312275299768, |
|
"eval_stsb_indo": 0.32549874112208793, |
|
"eval_stsb_spearman": 0.6668717733825881, |
|
"step": 520 |
|
}, |
|
{ |
|
"epoch": 2.14, |
|
"eval_CR": 72.09, |
|
"eval_FakeCLSDev": 0.9809523809523809, |
|
"eval_FakeCLSTest": 0.2903225806451613, |
|
"eval_FakeCLSTrain": 0.5580110497237569, |
|
"eval_FakePairs": -0.050563552137727515, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.54, |
|
"eval_MR": 63.24, |
|
"eval_MRPC": 70.68, |
|
"eval_SST2": 70.41, |
|
"eval_SUBJ": 83.0, |
|
"eval_TREC": 62.14, |
|
"eval_avg_sts": 0.6184279691881094, |
|
"eval_avg_transfer": 71.3, |
|
"eval_sickr_spearman": 0.5744775257457924, |
|
"eval_stsb_indo": 0.32854134925494444, |
|
"eval_stsb_spearman": 0.6623784126304263, |
|
"step": 540 |
|
}, |
|
{ |
|
"epoch": 2.22, |
|
"eval_CR": 72.34, |
|
"eval_FakeCLSDev": 0.9838709677419355, |
|
"eval_FakeCLSTest": 0.2857142857142857, |
|
"eval_FakeCLSTrain": 0.5187793427230047, |
|
"eval_FakePairs": -0.056554494334140726, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.73, |
|
"eval_MR": 63.1, |
|
"eval_MRPC": 70.66, |
|
"eval_SST2": 70.3, |
|
"eval_SUBJ": 83.1, |
|
"eval_TREC": 62.11, |
|
"eval_avg_sts": 0.6199240402637818, |
|
"eval_avg_transfer": 71.33428571428571, |
|
"eval_sickr_spearman": 0.5761467506140712, |
|
"eval_stsb_indo": 0.33424233003408477, |
|
"eval_stsb_spearman": 0.6637013299134924, |
|
"step": 560 |
|
}, |
|
{ |
|
"epoch": 2.3, |
|
"eval_CR": 72.49, |
|
"eval_FakeCLSDev": 0.9806451612903225, |
|
"eval_FakeCLSTest": 0.3, |
|
"eval_FakeCLSTrain": 0.553133514986376, |
|
"eval_FakePairs": -0.04856657140558978, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.9, |
|
"eval_MR": 62.85, |
|
"eval_MRPC": 70.53, |
|
"eval_SST2": 70.07, |
|
"eval_SUBJ": 83.09, |
|
"eval_TREC": 62.14, |
|
"eval_avg_sts": 0.6214529155922321, |
|
"eval_avg_transfer": 71.2957142857143, |
|
"eval_sickr_spearman": 0.5773829751028853, |
|
"eval_stsb_indo": 0.3387023269485864, |
|
"eval_stsb_spearman": 0.6655228560815791, |
|
"step": 580 |
|
}, |
|
{ |
|
"epoch": 2.38, |
|
"eval_CR": 72.1, |
|
"eval_FakeCLSDev": 0.9838056680161943, |
|
"eval_FakeCLSTest": 0.30857142857142855, |
|
"eval_FakeCLSTrain": 0.525, |
|
"eval_FakePairs": -0.04832693371773324, |
|
"eval_FakePairsNLI": 60.69, |
|
"eval_MPQA": 77.61, |
|
"eval_MR": 63.27, |
|
"eval_MRPC": 70.66, |
|
"eval_SST2": 70.07, |
|
"eval_SUBJ": 82.87, |
|
"eval_TREC": 62.12, |
|
"eval_avg_sts": 0.6206915906216466, |
|
"eval_avg_transfer": 71.24285714285715, |
|
"eval_sickr_spearman": 0.5774246660989336, |
|
"eval_stsb_indo": 0.33551007709822284, |
|
"eval_stsb_spearman": 0.6639585151443594, |
|
"step": 600 |
|
}, |
|
{ |
|
"epoch": 2.46, |
|
"eval_CR": 72.62, |
|
"eval_FakeCLSDev": 0.9790575916230366, |
|
"eval_FakeCLSTest": 0.3, |
|
"eval_FakeCLSTrain": 0.5488126649076517, |
|
"eval_FakePairs": -0.04353417996060268, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.77, |
|
"eval_MR": 62.79, |
|
"eval_MRPC": 70.09, |
|
"eval_SST2": 70.07, |
|
"eval_SUBJ": 82.63, |
|
"eval_TREC": 62.73, |
|
"eval_avg_sts": 0.6199540161828849, |
|
"eval_avg_transfer": 71.24285714285715, |
|
"eval_sickr_spearman": 0.5757875260064156, |
|
"eval_stsb_indo": 0.33257213690590665, |
|
"eval_stsb_spearman": 0.6641205063593542, |
|
"step": 620 |
|
}, |
|
{ |
|
"epoch": 2.54, |
|
"eval_CR": 72.58, |
|
"eval_FakeCLSDev": 0.9788359788359788, |
|
"eval_FakeCLSTest": 0.32, |
|
"eval_FakeCLSTrain": 0.5694050991501416, |
|
"eval_FakePairs": -0.03690420392990539, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.76, |
|
"eval_MR": 63.21, |
|
"eval_MRPC": 70.39, |
|
"eval_SST2": 69.61, |
|
"eval_SUBJ": 82.64, |
|
"eval_TREC": 62.91, |
|
"eval_avg_sts": 0.6197236868817497, |
|
"eval_avg_transfer": 71.3, |
|
"eval_sickr_spearman": 0.575295111154472, |
|
"eval_stsb_indo": 0.33277347755938264, |
|
"eval_stsb_spearman": 0.6641522626090275, |
|
"step": 640 |
|
}, |
|
{ |
|
"epoch": 2.62, |
|
"eval_CR": 72.37, |
|
"eval_FakeCLSDev": 0.9813664596273292, |
|
"eval_FakeCLSTest": 0.30714285714285716, |
|
"eval_FakeCLSTrain": 0.5673352435530086, |
|
"eval_FakePairs": -0.04185671614560698, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.7, |
|
"eval_MR": 63.44, |
|
"eval_MRPC": 70.41, |
|
"eval_SST2": 70.07, |
|
"eval_SUBJ": 82.42, |
|
"eval_TREC": 62.01, |
|
"eval_avg_sts": 0.6192603163969024, |
|
"eval_avg_transfer": 71.20285714285714, |
|
"eval_sickr_spearman": 0.5760667788301768, |
|
"eval_stsb_indo": 0.3335736566547829, |
|
"eval_stsb_spearman": 0.662453853963628, |
|
"step": 660 |
|
}, |
|
{ |
|
"epoch": 2.7, |
|
"eval_CR": 72.76, |
|
"eval_FakeCLSDev": 0.9801324503311258, |
|
"eval_FakeCLSTest": 0.31386861313868614, |
|
"eval_FakeCLSTrain": 0.5516304347826086, |
|
"eval_FakePairs": -0.037702996222760486, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.76, |
|
"eval_MR": 63.35, |
|
"eval_MRPC": 70.26, |
|
"eval_SST2": 69.72, |
|
"eval_SUBJ": 82.5, |
|
"eval_TREC": 62.16, |
|
"eval_avg_sts": 0.6187790665182422, |
|
"eval_avg_transfer": 71.21571428571428, |
|
"eval_sickr_spearman": 0.5762393065465424, |
|
"eval_stsb_indo": 0.33152148371183193, |
|
"eval_stsb_spearman": 0.661318826489942, |
|
"step": 680 |
|
}, |
|
{ |
|
"epoch": 2.78, |
|
"eval_CR": 72.7, |
|
"eval_FakeCLSDev": 0.9854014598540146, |
|
"eval_FakeCLSTest": 0.3153846153846154, |
|
"eval_FakeCLSTrain": 0.5739644970414202, |
|
"eval_FakePairs": -0.03586577394919376, |
|
"eval_FakePairsNLI": 61.15, |
|
"eval_MPQA": 77.57, |
|
"eval_MR": 63.55, |
|
"eval_MRPC": 70.49, |
|
"eval_SST2": 70.3, |
|
"eval_SUBJ": 82.3, |
|
"eval_TREC": 62.05, |
|
"eval_avg_sts": 0.6190232984727024, |
|
"eval_avg_transfer": 71.28, |
|
"eval_sickr_spearman": 0.5767789359711997, |
|
"eval_stsb_indo": 0.33084746387873076, |
|
"eval_stsb_spearman": 0.6612676609742051, |
|
"step": 700 |
|
}, |
|
{ |
|
"epoch": 2.86, |
|
"eval_CR": 72.72, |
|
"eval_FakeCLSDev": 0.9819277108433735, |
|
"eval_FakeCLSTest": 0.31690140845070425, |
|
"eval_FakeCLSTrain": 0.5483870967741935, |
|
"eval_FakePairs": -0.032590725548487874, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.62, |
|
"eval_MR": 63.28, |
|
"eval_MRPC": 70.22, |
|
"eval_SST2": 70.3, |
|
"eval_SUBJ": 82.42, |
|
"eval_TREC": 62.27, |
|
"eval_avg_sts": 0.619679371820255, |
|
"eval_avg_transfer": 71.26142857142858, |
|
"eval_sickr_spearman": 0.5773890270216666, |
|
"eval_stsb_indo": 0.3306356462208514, |
|
"eval_stsb_spearman": 0.6619697166188434, |
|
"step": 720 |
|
}, |
|
{ |
|
"epoch": 2.94, |
|
"eval_CR": 72.56, |
|
"eval_FakeCLSDev": 0.9830508474576272, |
|
"eval_FakeCLSTest": 0.31543624161073824, |
|
"eval_FakeCLSTrain": 0.5803571428571429, |
|
"eval_FakePairs": -0.03251084631920237, |
|
"eval_FakePairsNLI": 60.53, |
|
"eval_MPQA": 77.88, |
|
"eval_MR": 63.26, |
|
"eval_MRPC": 70.31, |
|
"eval_SST2": 70.3, |
|
"eval_SUBJ": 82.49, |
|
"eval_TREC": 62.01, |
|
"eval_avg_sts": 0.6197486139988778, |
|
"eval_avg_transfer": 71.25857142857143, |
|
"eval_sickr_spearman": 0.5774980576219311, |
|
"eval_stsb_indo": 0.3309848917313693, |
|
"eval_stsb_spearman": 0.6619991703758246, |
|
"step": 740 |
|
}, |
|
{ |
|
"epoch": 3.0, |
|
"step": 756, |
|
"train_runtime": 51381.4965, |
|
"train_samples_per_second": 0.015 |
|
} |
|
], |
|
"max_steps": 756, |
|
"num_train_epochs": 3, |
|
"total_flos": 24352519106985984, |
|
"trial_name": null, |
|
"trial_params": null |
|
} |
|
|