{ "model_dir": "models/prompt_tuning/llama-7b-code_easy", "distribution_id": "code_easy", "timestamp": "2023-10-19 22:48:34", "eval_loss": 0.2220458984375, "eval_score": -0.050140380859375, "eval_brier_score": 0.050140380859375, "eval_average_probability": 0.9130859375, "eval_accuracy": 0.9386666666666666, "eval_probabilities": [ 1.0, 0.97607421875, 1.0, 1.0, 1.0, 1.0, 0.262939453125, 1.0, 1.0, 0.9921875, 0.708984375, 0.99560546875, 0.998046875, 0.99951171875, 1.0, 0.95166015625, 0.99462890625, 1.0, 0.9775390625, 0.484375, 0.9599609375, 1.0, 1.0, 1.0, 1.0, 0.99951171875, 1.0, 0.9755859375, 1.0, 1.0, 0.93359375, 0.87646484375, 0.6376953125, 1.0, 0.0053863525390625, 0.265869140625, 0.77978515625, 1.0, 0.99853515625, 1.0, 1.0, 0.9150390625, 1.0, 0.880859375, 0.998046875, 0.00243377685546875, 0.1461181640625, 0.0004239082336425781, 0.8173828125, 0.93359375, 0.93701171875, 0.96142578125, 1.0, 0.99951171875, 1.0, 1.0, 1.0, 1.0, 0.69921875, 1.0, 0.99951171875, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99755859375, 0.00021827220916748047, 9.244680404663086e-05, 1.0, 1.0, 1.0, 0.99951171875, 0.59619140625, 1.0, 1.0, 0.94189453125, 1.0, 0.5849609375, 0.91259765625, 0.880859375, 1.0, 1.0, 1.0, 0.82080078125, 0.974609375, 0.99365234375, 1.0, 0.303955078125, 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, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9931640625, 0.80615234375, 1.0, 0.83544921875, 1.0, 1.0, 0.99951171875, 1.0, 1.0, 1.0, 1.0, 1.0, 0.97412109375, 1.0, 0.880859375, 1.0, 0.98876953125, 1.0, 1.0, 1.0, 1.0, 0.7548828125, 1.0, 0.998046875, 0.990234375, 0.99951171875, 0.99951171875, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9228515625, 0.359375, 0.95458984375, 0.9931640625, 1.0, 0.97998046875, 1.0, 1.0, 1.0, 0.89013671875, 0.921875, 0.73095703125, 0.9990234375, 0.5390625, 0.9599609375, 0.822265625, 0.60009765625, 0.85791015625, 0.9619140625, 0.99853515625, 1.0, 1.0, 1.0, 0.99755859375, 1.0, 0.9677734375, 0.99267578125, 1.0, 1.0, 1.0, 0.98193359375, 1.0, 0.99951171875, 1.0, 0.99658203125, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.96484375, 1.0, 1.0, 1.0, 0.880859375, 1.0, 0.99560546875, 0.96337890625, 0.99951171875, 0.71533203125, 0.86328125, 0.955078125, 0.978515625, 0.986328125, 0.52734375, 0.57763671875, 0.69287109375, 0.99951171875, 1.0, 0.96044921875, 1.0, 0.986328125, 1.0, 1.0, 1.0, 0.9326171875, 1.0, 0.99951171875, 1.0, 0.99951171875, 1.0, 0.9794921875, 1.0, 1.0, 1.0, 0.9970703125, 0.998046875, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.7548828125, 0.99755859375, 0.89306640625, 6.604194641113281e-05, 1.0, 1.0, 0.9658203125, 1.0, 0.92626953125, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.88427734375, 0.9638671875, 0.5234375, 0.7216796875, 1.0, 1.0, 1.0, 0.865234375, 0.78515625, 0.77587890625, 0.96533203125, 0.85986328125, 0.91259765625, 1.0, 1.0, 1.0, 0.5849609375, 1.0, 0.658203125, 1.0, 0.99951171875, 0.85791015625, 1.0, 1.0, 1.0, 0.5068359375, 0.1865234375, 0.51123046875, 0.445556640625, 0.9365234375, 1.0, 1.0, 1.0, 1.0, 0.6513671875, 1.0, 0.77978515625, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.89501953125, 0.94970703125, 0.98681640625, 1.0, 0.9375, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99560546875, 0.98681640625, 1.0, 1.0, 1.0, 0.998046875, 1.0, 1.0, 1.0, 0.94970703125, 1.0, 1.0, 0.77197265625, 1.0, 1.0, 0.70068359375, 1.0, 1.0, 1.0, 1.0, 0.99951171875, 0.8017578125, 0.939453125, 0.57763671875, 0.443603515625, 0.54296875, 1.0, 1.0, 0.95263671875, 1.0, 1.0, 1.0, 0.99609375, 1.0, 1.0, 0.958984375, 1.0, 0.77490234375, 0.99658203125, 1.0, 1.0, 1.0, 1.0, 1.0, 0.97119140625, 0.9677734375, 0.99951171875, 1.0, 1.0, 1.0, 0.83984375, 1.0, 1.0, 0.986328125, 1.0, 0.99951171875, 1.0, 0.388671875, 0.8798828125, 1.0, 0.99267578125, 0.98486328125, 1.0, 1.0, 0.7490234375, 0.99853515625, 0.896484375, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99951171875, 1.0, 1.0, 0.99951171875, 1.0, 1.0, 1.0, 1.0, 0.32080078125, 0.01363372802734375, 0.83349609375, 0.97509765625, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.7548828125, 1.0, 0.9990234375, 1.0, 1.0, 1.0, 0.99951171875, 0.3486328125, 1.0, 1.0, 1.0, 1.0, 0.99951171875, 1.0, 1.0, 0.8759765625, 0.88623046875, 0.74853515625, 1.0, 1.0, 1.0, 0.98095703125, 1.0, 0.99951171875, 0.219970703125, 0.56982421875, 0.71240234375, 0.986328125, 1.0, 0.89013671875, 0.99951171875, 1.0, 0.99951171875, 1.0, 1.0, 1.0, 1.0, 0.98779296875, 0.998046875, 1.0, 1.0, 0.99853515625, 1.0, 1.0, 1.0, 1.0, 1.0, 0.89306640625, 1.0, 1.0, 0.9990234375, 0.615234375, 0.58154296875, 0.43017578125, 0.97705078125, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.94970703125, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99072265625, 0.99072265625, 0.262939453125, 0.92626953125, 0.9150390625, 0.99853515625, 5.662441253662109e-06, 0.235107421875, 0.9970703125, 0.9912109375, 0.5908203125, 0.0010166168212890625, 0.81884765625, 0.1993408203125, 1.0, 1.0, 1.0, 0.98779296875, 0.058349609375, 1.0, 0.99951171875, 1.0, 1.0, 0.615234375, 1.0, 0.79296875, 1.0, 1.0, 0.9970703125, 0.62255859375, 0.67919921875, 0.93994140625, 1.0, 0.9306640625, 1.0, 0.93798828125, 0.99951171875, 0.99951171875, 0.99951171875, 1.0, 1.0, 1.0, 0.99951171875, 0.8173828125, 0.99853515625, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.998046875, 0.658203125, 0.99609375, 0.5, 1.0, 1.0, 1.0, 1.0, 0.994140625, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99951171875, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99609375, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99951171875, 1.0, 0.9990234375, 1.0, 1.0, 0.08880615234375, 1.0, 1.0, 1.0, 1.0, 1.0, 0.99462890625, 0.8916015625, 0.99072265625, 0.94482421875, 1.0, 0.99658203125, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.06610107421875, 1.0, 1.0, 0.32763671875, 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, 0.99755859375, 0.97802734375, 0.60400390625, 0.98486328125, 0.77197265625, 0.58154296875, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.9990234375, 0.98388671875, 0.94970703125, 0.32080078125, 0.029754638671875, 1.0, 0.99951171875, 1.0, 1.0, 1.0, 1.0, 1.0, 0.615234375, 0.5888671875, 1.0, 1.0, 0.99951171875, 0.9482421875, 1.0, 1.0, 1.0, 0.57373046875, 0.7431640625, 0.56201171875, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.96875, 0.982421875, 0.822265625, 0.8759765625, 0.9677734375, 0.8173828125, 1.0, 1.0, 0.99560546875, 1.0, 1.0, 1.0, 0.64794921875, 0.2568359375, 0.2147216796875, 1.0, 0.56982421875, 1.0, 1.0, 1.0, 1.0, 1.0, 0.85986328125, 1.0, 0.007122039794921875, 0.83740234375, 0.2017822265625, 0.9990234375, 0.99951171875, 0.99755859375, 1.0, 0.26904296875, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 0.56201171875, 0.921875, 0.9794921875, 0.60009765625, 0.98291015625, 0.4609375, 1.0, 1.0, 1.0, 0.99169921875, 0.99560546875, 0.93701171875, 1.0, 1.0, 1.0, 0.96630859375, 0.9501953125, 0.9150390625, 0.99560546875, 1.0, 0.98486328125, 0.841796875, 0.61865234375, 0.998046875, 1.0, 0.99951171875, 0.97998046875, 1.0, 1.0, 1.0, 1.0, 0.98779296875, 1.0, 0.99755859375, 0.98876953125, 0.8740234375, 0.9814453125, 0.90185546875, 0.34521484375, 0.0009398460388183594, 1.0, 0.43017578125, 1.0, 1.0, 1.0, 0.29736328125, 0.974609375, 0.19677734375, 0.359375, 0.373779296875, 0.49609375, 1.0, 1.0, 1.0, 0.9814453125, 0.994140625, 0.93359375, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 ], "eval_runtime": 13.0708, "eval_samples_per_second": 57.38, "eval_steps_per_second": 1.836 }