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

765 lines
13 KiB
JSON

{
"model_dir": "models/prompt_tuning/llama-30b-code_easy",
"distribution_id": "code_easy",
"timestamp": "2023-10-13 17:22:09",
"eval_loss": 0.1318359375,
"eval_score": -0.0352783203125,
"eval_brier_score": 0.0352783203125,
"eval_average_probability": 0.9306640625,
"eval_accuracy": 0.96,
"eval_probabilities": [
0.998046875,
1.0,
1.0,
1.0,
0.9990234375,
0.99951171875,
0.53125,
1.0,
1.0,
0.8701171875,
0.62353515625,
0.9228515625,
0.90576171875,
1.0,
0.99951171875,
0.89892578125,
1.0,
1.0,
0.974609375,
0.94873046875,
0.74755859375,
0.99951171875,
1.0,
1.0,
1.0,
0.99951171875,
0.99951171875,
1.0,
1.0,
1.0,
0.97021484375,
0.99267578125,
0.986328125,
0.93994140625,
0.9765625,
0.69921875,
0.73583984375,
1.0,
0.9873046875,
0.99755859375,
1.0,
0.994140625,
1.0,
1.0,
0.99951171875,
0.9912109375,
0.71875,
0.9423828125,
0.7880859375,
1.0,
0.99462890625,
0.5390625,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9404296875,
0.662109375,
0.9921875,
0.97509765625,
1.0,
0.99853515625,
1.0,
1.0,
1.0,
0.99951171875,
0.99951171875,
0.9990234375,
0.65087890625,
0.0001308917999267578,
0.00217437744140625,
1.0,
1.0,
0.9970703125,
0.9775390625,
0.94482421875,
1.0,
1.0,
0.998046875,
0.99951171875,
0.87255859375,
0.98779296875,
0.7373046875,
0.9892578125,
1.0,
0.99951171875,
0.99951171875,
0.8681640625,
0.9736328125,
1.0,
0.98974609375,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.96484375,
0.99853515625,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.5849609375,
1.0,
0.392333984375,
1.0,
0.99560546875,
0.90283203125,
0.998046875,
0.99560546875,
1.0,
1.0,
1.0,
0.7490234375,
0.99951171875,
0.98681640625,
0.9990234375,
0.99560546875,
0.994140625,
1.0,
1.0,
1.0,
0.52734375,
1.0,
0.0994873046875,
0.79833984375,
1.0,
0.994140625,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.99609375,
1.0,
1.0,
0.9111328125,
0.515625,
0.97509765625,
0.9658203125,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.7431640625,
0.978515625,
0.4921875,
0.9501953125,
0.79833984375,
0.71875,
0.640625,
0.615234375,
1.0,
0.99267578125,
1.0,
1.0,
1.0,
0.98779296875,
0.9970703125,
0.99560546875,
0.98291015625,
1.0,
1.0,
1.0,
0.7607421875,
1.0,
0.9970703125,
1.0,
0.9990234375,
1.0,
1.0,
1.0,
1.0,
0.99951171875,
1.0,
1.0,
1.0,
0.87744140625,
1.0,
1.0,
0.99853515625,
0.98291015625,
1.0,
0.9990234375,
0.998046875,
1.0,
0.740234375,
0.99853515625,
0.9931640625,
0.2509765625,
1.0,
0.68408203125,
0.8642578125,
0.8662109375,
1.0,
0.9365234375,
0.99755859375,
1.0,
0.99658203125,
1.0,
0.9990234375,
0.99951171875,
0.99951171875,
1.0,
1.0,
1.0,
0.99951171875,
1.0,
1.0,
1.0,
1.0,
1.0,
0.98876953125,
0.9990234375,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.951171875,
0.70263671875,
1.0,
0.9970703125,
0.1441650390625,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9990234375,
0.88720703125,
0.437744140625,
0.338134765625,
1.0,
1.0,
1.0,
0.8515625,
0.7890625,
0.67236328125,
0.99365234375,
0.9951171875,
0.998046875,
1.0,
1.0,
0.248046875,
0.67919921875,
1.0,
0.77978515625,
0.99951171875,
0.99951171875,
0.99951171875,
1.0,
1.0,
0.99951171875,
0.5830078125,
0.9375,
0.6279296875,
0.67236328125,
0.99365234375,
0.9990234375,
1.0,
1.0,
1.0,
0.68603515625,
0.99951171875,
0.77197265625,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.94580078125,
0.99951171875,
0.998046875,
1.0,
0.9111328125,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.68603515625,
1.0,
1.0,
1.0,
0.98876953125,
1.0,
1.0,
1.0,
1.0,
0.403564453125,
0.97509765625,
1.0,
1.0,
1.0,
1.0,
0.99951171875,
0.99609375,
0.97119140625,
0.962890625,
0.94970703125,
0.83935546875,
1.0,
1.0,
0.73388671875,
1.0,
1.0,
0.99951171875,
1.0,
1.0,
1.0,
1.0,
1.0,
0.0843505859375,
0.99951171875,
1.0,
1.0,
1.0,
1.0,
1.0,
0.98779296875,
0.7373046875,
0.9990234375,
1.0,
1.0,
1.0,
0.8662109375,
1.0,
1.0,
0.9990234375,
0.998046875,
0.990234375,
1.0,
0.70263671875,
0.873046875,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.98876953125,
0.955078125,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9970703125,
0.352294921875,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.810546875,
0.0016870498657226562,
0.81884765625,
0.9990234375,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.955078125,
1.0,
1.0,
1.0,
1.0,
1.0,
0.99951171875,
0.9970703125,
1.0,
0.994140625,
1.0,
0.966796875,
1.0,
1.0,
0.99951171875,
0.841796875,
0.69921875,
0.822265625,
1.0,
1.0,
1.0,
0.96435546875,
0.9990234375,
0.9951171875,
0.060089111328125,
0.09600830078125,
0.09271240234375,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.98046875,
0.99853515625,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.97998046875,
1.0,
0.9658203125,
0.2783203125,
0.990234375,
0.81982421875,
0.62255859375,
0.998046875,
1.0,
1.0,
1.0,
1.0,
0.9921875,
1.0,
0.99853515625,
1.0,
0.880859375,
1.0,
1.0,
1.0,
0.99951171875,
0.99951171875,
1.0,
1.0,
1.0,
0.978515625,
1.0,
0.79296875,
0.57763671875,
1.0,
0.99658203125,
1.0,
0.615234375,
0.69287109375,
0.9697265625,
1.0,
0.662109375,
0.96484375,
0.76806640625,
0.8173828125,
1.0,
1.0,
1.0,
0.9873046875,
0.9736328125,
0.9990234375,
0.99853515625,
1.0,
1.0,
0.7373046875,
1.0,
0.120849609375,
1.0,
1.0,
0.9599609375,
0.310791015625,
0.8759765625,
0.97216796875,
1.0,
0.85986328125,
1.0,
0.96044921875,
1.0,
1.0,
1.0,
0.98779296875,
1.0,
1.0,
1.0,
0.607421875,
0.9814453125,
0.994140625,
0.99951171875,
1.0,
0.99951171875,
1.0,
1.0,
0.99951171875,
0.99853515625,
0.751953125,
0.99951171875,
0.98681640625,
1.0,
1.0,
1.0,
1.0,
0.96484375,
1.0,
1.0,
1.0,
1.0,
1.0,
0.99951171875,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.98974609375,
1.0,
1.0,
0.95458984375,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9990234375,
1.0,
0.970703125,
1.0,
1.0,
1.0,
1.0,
0.9921875,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.58154296875,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.998046875,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9990234375,
0.9990234375,
0.982421875,
0.9833984375,
0.012237548828125,
0.99951171875,
0.9990234375,
0.99755859375,
1.0,
1.0,
1.0,
0.484375,
0.73095703125,
0.53515625,
1.0,
0.1312255859375,
1.0,
0.98095703125,
0.97119140625,
1.0,
1.0,
1.0,
0.9951171875,
1.0,
0.9755859375,
1.0,
0.994140625,
0.986328125,
0.689453125,
1.0,
1.0,
0.98095703125,
0.453369140625,
0.88427734375,
0.6298828125,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.9853515625,
0.90478515625,
0.615234375,
0.607421875,
0.6826171875,
0.87060546875,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.951171875,
0.98193359375,
0.8154296875,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.96875,
1.0,
0.94580078125,
0.11199951171875,
0.20556640625,
0.9990234375,
1.0,
1.0,
1.0,
0.8154296875,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
0.99853515625,
1.0,
1.0,
1.0,
0.94677734375,
0.552734375,
0.99951171875,
1.0,
1.0,
0.5546875,
0.5546875,
0.98828125,
1.0,
1.0,
1.0,
0.99658203125,
0.5927734375,
0.291015625,
1.0,
1.0,
0.9833984375,
0.85791015625,
0.64794921875,
0.9794921875,
1.0,
0.865234375,
0.97998046875,
1.0,
1.0,
1.0,
1.0,
0.9912109375,
1.0,
0.8857421875,
0.921875,
0.56982421875,
0.95849609375,
0.9990234375,
0.91845703125,
0.449462890625,
1.0,
0.90869140625,
1.0,
1.0,
0.99853515625,
0.99951171875,
1.0,
0.403564453125,
0.9111328125,
0.363037109375,
0.69287109375,
1.0,
1.0,
1.0,
0.70556640625,
0.90185546875,
0.8916015625,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0
],
"eval_runtime": 31.6859,
"eval_samples_per_second": 23.67,
"eval_steps_per_second": 0.757
}