JavaBench / data /data_context.json
CPunisher's picture
Data
23f22ff
{
"completion": [
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-4o-2024-05-13",
"Pass_at_1": 1.0
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.9714285714
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.945
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.9378571429
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.9357142857
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.9328571429
},
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.9214285714
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.9007142857
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.8907142857
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.8828571429
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.8728571429
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.8578571429
},
{
"Context": "selective",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.845
},
{
"Context": "selective",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.7964285714
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.7935714286
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.7192857143
}
],
"compilation_class_wise": [
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.7942857143
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.7414285714
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.7385714286
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.7314285714
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.7171428571
},
{
"Context": "selective",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.6978571429
},
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-4o-2024-05-13",
"Pass_at_1": 0.6607142857
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.6592857143
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.6414285714
},
{
"Context": "selective",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.58
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.4814285714
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.45
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.39
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.3692857143
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.3457142857
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.215
}
],
"compilation_test_wise": [
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-4o-2024-05-13",
"Pass_at_1": 0.5035714286
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.4202826585
},
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.3443277311
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.3405987395
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.3387079832
},
{
"Context": "selective",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.3183823529
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.3121848739
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.2858193277
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.2283088235
},
{
"Context": "selective",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.2240546218
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.1466911765
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.1128676471
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.03125
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.0147058824
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.0125
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.0
}
],
"pass_class_wise": [
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.7832360347
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.723699056
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.715291943
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.7033228696
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.6855203826
},
{
"Context": "selective",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.6808480861
},
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-4o-2024-05-13",
"Pass_at_1": 0.6545897285
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.6417690022
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.6293667264
},
{
"Context": "selective",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.5674101922
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.4741970721
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.4489821662
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.3864123669
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.3599216366
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.3272885064
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.2139054163
}
],
"pass_test_wise": [
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-4o-2024-05-13",
"Pass_at_1": 0.3438179726
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.3047552867
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.3032497787
},
{
"Context": "selective",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.2941156144
},
{
"Context": "selective",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.2544265255
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.2393000344
},
{
"Context": "selective",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.2028454735
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.1966660863
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.1877858469
},
{
"Context": "selective",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.1669267449
},
{
"Context": "maximum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.1370849195
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-33b-instruct",
"Pass_at_1": 0.1123372221
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "gpt-3.5-turbo-1106",
"Pass_at_1": 0.03125
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "WizardCoder-15B-V1.0",
"Pass_at_1": 0.0147058824
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "deepseek-coder-6.7b-instruct",
"Pass_at_1": 0.0125
},
{
"Context": "minimum",
"Method": "holistic",
"Model": "Phind-CodeLlama-34B-v2",
"Pass_at_1": 0.0
}
]
}