Files
GENIES/results/evaluations/prompt_tuning/llama-30b-code_easy/code_hard.json
T
2023-11-11 19:44:09 +00:00

766 lines
17 KiB
JSON

{
"model_dir": "models/prompt_tuning/llama-30b-code_easy",
"distribution_id": "code_hard",
"timestamp": "2023-10-13 17:20:15",
"eval_loss": 1.50390625,
"eval_score": -0.293212890625,
"eval_brier_score": 0.2931969475294972,
"eval_average_probability": 0.5739171963532765,
"eval_accuracy": 0.6493333333333333,
"eval_probabilities": [
0.9990234375,
0.39306640625,
0.56982421875,
0.99853515625,
0.99755859375,
0.9833984375,
0.607421875,
0.51171875,
0.08331298828125,
0.2451171875,
0.1961669921875,
0.2958984375,
0.9921875,
4.6253204345703125e-05,
1.1146068572998047e-05,
0.7626953125,
0.802734375,
0.8544921875,
0.6875,
0.99658203125,
0.9609375,
0.66552734375,
0.310791015625,
0.83349609375,
0.763671875,
0.82421875,
0.9658203125,
0.9638671875,
0.59326171875,
0.8955078125,
0.751953125,
0.71533203125,
0.64404296875,
0.289306640625,
0.053009033203125,
0.07208251953125,
0.5,
0.513671875,
1.0,
0.943359375,
0.8837890625,
0.99560546875,
0.2186279296875,
0.138427734375,
0.158203125,
0.99267578125,
0.517578125,
0.501953125,
0.72412109375,
0.99951171875,
0.99951171875,
0.359375,
0.007232666015625,
0.004230499267578125,
0.00016832351684570312,
0.00010699033737182617,
3.993511199951172e-06,
0.60009765625,
0.505859375,
0.355712890625,
7.271766662597656e-06,
8.165836334228516e-06,
8.761882781982422e-06,
0.015838623046875,
0.52099609375,
0.98974609375,
0.199462890625,
0.966796875,
0.1514892578125,
0.97705078125,
0.9921875,
0.96826171875,
0.953125,
0.79833984375,
0.72509765625,
0.9794921875,
0.73388671875,
0.67041015625,
0.56591796875,
0.658203125,
0.438720703125,
0.501953125,
0.50390625,
0.62060546875,
0.71875,
0.95849609375,
0.611328125,
0.43017578125,
0.5234375,
0.56982421875,
0.68994140625,
0.94970703125,
0.8984375,
1.0,
0.99951171875,
1.0,
0.98046875,
0.9951171875,
0.822265625,
0.9951171875,
0.9833984375,
0.943359375,
0.90869140625,
0.60009765625,
0.60205078125,
0.0869140625,
0.99609375,
0.1346435546875,
0.1441650390625,
0.498046875,
0.482421875,
2.658367156982422e-05,
0.00327301025390625,
0.00511932373046875,
0.75341796875,
0.26123046875,
0.642578125,
0.19189453125,
0.294677734375,
0.60302734375,
0.75048828125,
0.873046875,
0.375244140625,
0.55029296875,
0.59619140625,
1.0,
0.9755859375,
0.82958984375,
0.1441650390625,
0.355712890625,
0.050323486328125,
0.0330810546875,
0.1431884765625,
0.1710205078125,
0.19189453125,
0.99755859375,
0.98974609375,
0.998046875,
0.033538818359375,
0.04876708984375,
0.04949951171875,
0.082275390625,
0.055511474609375,
0.0517578125,
1.0,
0.67919921875,
0.5078125,
0.0001310110092163086,
0.005641937255859375,
0.0074615478515625,
0.76611328125,
0.87255859375,
0.99951171875,
0.363037109375,
0.036773681640625,
0.04534912109375,
0.65478515625,
0.90966796875,
0.6435546875,
0.95751953125,
0.99853515625,
0.962890625,
0.69677734375,
0.548828125,
0.5947265625,
0.466796875,
0.5869140625,
0.541015625,
0.658203125,
0.6123046875,
0.55908203125,
0.9736328125,
0.49609375,
0.63330078125,
0.99609375,
0.85498046875,
0.92919921875,
0.94482421875,
0.1859130859375,
0.029541015625,
0.841796875,
0.7880859375,
0.640625,
1.0,
0.51171875,
0.810546875,
0.67041015625,
0.83251953125,
0.8330078125,
0.75634765625,
0.85888671875,
0.64404296875,
0.5234375,
0.60009765625,
0.46484375,
3.2782554626464844e-06,
5.304813385009766e-06,
5.960464477539063e-08,
0.61865234375,
0.333984375,
0.2109375,
1.5497207641601562e-06,
5.841255187988281e-06,
4.5299530029296875e-06,
0.5078125,
0.59814453125,
0.5,
0.00278472900390625,
0.0014925003051757812,
0.0007395744323730469,
1.0,
1.0,
1.0,
0.466796875,
0.484375,
0.4765625,
0.63720703125,
0.6298828125,
0.615234375,
0.6240234375,
0.4853515625,
0.62841796875,
0.4912109375,
0.27978515625,
0.4609375,
0.5830078125,
0.55810546875,
1.0,
0.6689453125,
0.37353515625,
0.7939453125,
0.58984375,
0.677734375,
0.424072265625,
0.8251953125,
0.82373046875,
0.72314453125,
0.99951171875,
0.99560546875,
0.98681640625,
0.896484375,
0.98779296875,
0.5849609375,
0.4072265625,
0.85498046875,
0.4609375,
0.60107421875,
0.57470703125,
0.548828125,
0.61865234375,
0.56005859375,
0.66552734375,
0.509765625,
0.48583984375,
0.438720703125,
0.0160675048828125,
0.0255584716796875,
0.00981903076171875,
0.65283203125,
0.8642578125,
0.46728515625,
0.91943359375,
0.90185546875,
0.97216796875,
0.6875,
0.767578125,
0.69287109375,
0.732421875,
0.7578125,
0.8310546875,
0.99755859375,
1.0,
1.0,
0.67236328125,
0.60009765625,
0.8232421875,
1.0,
1.0,
1.0,
0.71533203125,
0.5947265625,
0.810546875,
0.96630859375,
0.84814453125,
0.9404296875,
0.26904296875,
0.890625,
0.452392578125,
0.98876953125,
0.921875,
0.97265625,
0.51953125,
0.60205078125,
0.57373046875,
0.012237548828125,
0.01238250732421875,
0.440673828125,
0.490234375,
0.60107421875,
0.505859375,
0.87890625,
0.9541015625,
0.900390625,
0.355712890625,
0.7431640625,
0.98291015625,
0.00012481212615966797,
0.00011366605758666992,
1.0,
0.30078125,
0.78515625,
0.401611328125,
0.8740234375,
0.53515625,
0.56201171875,
0.5966796875,
0.30908203125,
0.06402587890625,
1.0,
0.355712890625,
0.79296875,
0.60693359375,
0.5341796875,
0.548828125,
0.986328125,
0.0008945465087890625,
0.0006341934204101562,
0.611328125,
0.97216796875,
1.0,
0.39990234375,
0.63232421875,
0.404541015625,
0.9990234375,
0.01509857177734375,
0.03521728515625,
0.56103515625,
0.548828125,
0.49609375,
0.53173828125,
0.43359375,
0.52783203125,
0.51171875,
0.74609375,
0.87060546875,
5.805492401123047e-05,
6.181001663208008e-05,
0.00021827220916748047,
0.96875,
0.94287109375,
0.970703125,
0.89501953125,
0.32763671875,
0.81298828125,
0.0047607421875,
0.00262451171875,
0.0017423629760742188,
0.9580078125,
0.96826171875,
0.95556640625,
0.97998046875,
0.98583984375,
0.9794921875,
0.9912109375,
0.5927734375,
0.505859375,
0.97412109375,
0.95263671875,
0.9619140625,
0.67724609375,
0.6669921875,
0.66162109375,
0.031982421875,
0.51220703125,
0.428466796875,
0.70556640625,
0.0007138252258300781,
0.001953125,
0.0010862350463867188,
0.0024242401123046875,
0.0162200927734375,
0.67578125,
0.315673828125,
0.1561279296875,
0.047607421875,
0.260498046875,
0.505859375,
0.494140625,
0.373779296875,
0.5,
0.884765625,
0.85888671875,
0.9736328125,
0.418701171875,
0.46875,
0.60009765625,
0.64599609375,
0.515625,
0.86865234375,
0.2281494140625,
0.51171875,
0.9619140625,
0.07208251953125,
0.1500244140625,
1.0,
0.94921875,
0.99951171875,
0.75439453125,
0.369873046875,
0.248291015625,
0.51123046875,
0.84814453125,
0.85986328125,
0.80810546875,
0.4921875,
0.56591796875,
0.307373046875,
0.82861328125,
0.98095703125,
0.68359375,
0.9365234375,
0.93505859375,
0.90283203125,
0.154052734375,
0.0009698867797851562,
0.0007791519165039062,
0.76708984375,
0.54345703125,
0.59912109375,
7.086992263793945e-05,
2.5451183319091797e-05,
1.4901161193847656e-06,
0.5234375,
0.7958984375,
0.8623046875,
0.262939453125,
0.3173828125,
0.7841796875,
0.7099609375,
0.0325927734375,
0.001956939697265625,
0.5458984375,
0.51220703125,
0.5908203125,
0.517578125,
0.548828125,
0.5517578125,
0.93017578125,
0.437744140625,
0.689453125,
0.355712890625,
0.24365234375,
0.3486328125,
0.689453125,
0.57568359375,
0.8984375,
0.53125,
0.51171875,
0.51953125,
0.53515625,
0.7890625,
0.5234375,
0.54296875,
0.638671875,
0.1993408203125,
0.7607421875,
0.8291015625,
0.6826171875,
0.77685546875,
0.99853515625,
0.89306640625,
3.88026237487793e-05,
5.9604644775390625e-06,
3.7550926208496094e-06,
0.11041259765625,
0.146240234375,
0.99853515625,
0.91943359375,
0.54296875,
0.484375,
0.312744140625,
0.53564453125,
0.82861328125,
0.3251953125,
0.56494140625,
0.27587890625,
0.9658203125,
0.9755859375,
0.77783203125,
0.8310546875,
0.2421875,
0.82666015625,
0.607421875,
0.57958984375,
0.1175537109375,
0.732421875,
0.916015625,
0.8740234375,
0.92138671875,
0.9287109375,
0.896484375,
0.0775146484375,
0.77734375,
0.9814453125,
0.90283203125,
0.1365966796875,
0.0247955322265625,
5.7220458984375e-06,
2.390146255493164e-05,
5.918741226196289e-05,
0.9892578125,
0.9794921875,
0.449462890625,
0.962890625,
0.6298828125,
0.307373046875,
0.99853515625,
0.99951171875,
0.99853515625,
0.9619140625,
0.12335205078125,
0.06805419921875,
0.90380859375,
0.85595703125,
0.830078125,
2.2470951080322266e-05,
2.1755695343017578e-05,
0.0010986328125,
0.9111328125,
0.90869140625,
0.89501953125,
0.693359375,
0.06097412109375,
0.023193359375,
4.57763671875e-05,
2.9087066650390625e-05,
3.3974647521972656e-05,
0.99658203125,
0.5830078125,
0.414794921875,
0.927734375,
0.83251953125,
0.900390625,
0.1900634765625,
0.369384765625,
0.342041015625,
0.732421875,
0.66357421875,
0.642578125,
0.994140625,
0.7880859375,
0.9755859375,
0.515625,
0.59814453125,
0.9111328125,
0.1973876953125,
0.2421875,
0.23095703125,
1.0,
0.56591796875,
0.9345703125,
0.7890625,
0.94287109375,
0.892578125,
0.1331787109375,
0.31005859375,
0.1815185546875,
0.97265625,
0.5,
1.0,
0.98046875,
0.8203125,
0.88623046875,
0.00821685791015625,
0.005321502685546875,
0.0035247802734375,
0.5546875,
0.51171875,
0.48046875,
0.7548828125,
1.0,
0.6298828125,
0.00278472900390625,
1.0,
0.0206451416015625,
0.07696533203125,
0.03759765625,
0.08758544921875,
3.272294998168945e-05,
4.1961669921875e-05,
3.069639205932617e-05,
0.5322265625,
0.406494140625,
0.63623046875,
0.437744140625,
0.5224609375,
0.56005859375,
0.71240234375,
0.658203125,
0.78271484375,
0.8291015625,
0.9853515625,
0.56884765625,
0.49072265625,
0.182373046875,
0.99462890625,
0.64453125,
0.95458984375,
0.98095703125,
0.462890625,
0.5849609375,
0.4716796875,
0.0244293212890625,
0.0247955322265625,
0.0236968994140625,
1.0,
1.0,
1.0,
0.4921875,
0.57373046875,
0.29736328125,
0.45751953125,
0.802734375,
0.54345703125,
0.84814453125,
0.449462890625,
0.96142578125,
1.0,
0.533203125,
0.4892578125,
0.521484375,
0.49609375,
0.64794921875,
0.036773681640625,
0.005641937255859375,
0.006439208984375,
0.5224609375,
0.53125,
0.43408203125,
0.420654296875,
0.501953125,
0.49609375,
0.57958984375,
0.96630859375,
0.99169921875,
0.76611328125,
0.8740234375,
0.5546875,
0.73486328125,
0.85693359375,
0.763671875,
0.70263671875,
0.896484375,
0.6826171875,
0.414794921875,
0.416748046875,
0.5546875,
0.275146484375,
0.35400390625,
0.20751953125,
0.693359375,
0.923828125,
0.96142578125,
0.88720703125,
0.8662109375,
0.82666015625,
0.642578125,
0.74755859375,
0.51171875,
0.6259765625,
0.72314453125,
0.79052734375,
0.6689453125,
0.57373046875,
0.6298828125,
0.5380859375,
0.51513671875,
0.2437744140625,
0.55810546875,
0.6298828125,
0.53125,
0.85986328125,
0.90185546875,
0.9599609375,
0.46484375,
0.5,
0.9921875,
0.06903076171875,
8.684396743774414e-05,
0.0074615478515625,
0.5546875,
0.56494140625,
0.900390625,
1.0,
1.0,
1.0,
0.348388671875,
0.65625,
0.8095703125,
6.949901580810547e-05,
2.2590160369873047e-05,
2.3126602172851562e-05,
0.455078125,
0.0328369140625,
0.90087890625,
0.392333984375,
0.308837890625,
0.470703125,
0.294189453125,
0.6259765625,
0.275146484375,
0.9052734375,
0.86083984375,
0.923828125,
0.47265625,
0.5888671875,
0.77734375,
0.607421875,
0.67578125,
0.34521484375,
0.951171875,
0.9345703125,
0.84814453125,
1.0,
1.0,
1.0,
0.998046875,
0.99853515625,
0.99609375,
0.99951171875,
0.0002269744873046875,
0.00031256675720214844,
0.8642578125,
0.89013671875,
1.0,
0.357666015625,
0.0374755859375,
0.0305633544921875,
0.8818359375,
0.86328125,
0.87744140625,
0.11517333984375,
0.73193359375,
0.6689453125,
4.172325134277344e-07,
4.76837158203125e-07,
5.364418029785156e-07,
0.99951171875,
0.99951171875,
0.99951171875
],
"eval_runtime": 96.714,
"eval_samples_per_second": 7.755,
"eval_steps_per_second": 0.248,
"eval_rms_calibration_error": 0.3045070931297283
}