IC-1413GPT-4 achieves near-saturation on Python code synthesis (86.6% pass@1) but scores significantly lower on code repair (47.8% avg) and code explanation (52.1% avg) across six languages

Niklas Muennighoff, Qian Liu, Armel Randy Zebaze, Qinkai Zheng, Binyuan Hui, Terry Yue Zhuo, Swayam Singh, Xiangru Tang, Leandro Von Werra, Shayne Longpre

SourceOctoPack: Instruction Tuning Code Large Language Models

The paper evaluates GPT-4 on HumanEvalPack, a benchmark spanning three coding tasks (synthesis, repair, explanation) across six programming languages. On the synthesis task, GPT-4 achieves 86.6% pass@1 on Python, close to the maximum possible. However, on the repair and explanation tasks, GPT-4's average scores drop to 47.8% and 52.1% respectively. The authors note this suggests the original HumanEval benchmark may soon cease to be useful due to models reaching near-maximum performance on synthesis, and that the new repair and explanation variants provide room for future improvement.

Evidence
correlational
Key metric
86.6% pass@1 Python HumanEvalSynthesize; 47.8% avg HumanEvalFix; 52.1% avg HumanEvalExplain (6 languages)
Caveat
gpt-4 was originally found to score 67% on python humaneval (openai, 2023) and 81% in later work (bubeck et al., 2023). our score for gpt-4 is significantly higher, possibly due to improvements made to the api by openai, contamination of humaneval in gpt-4 training, or slightly different prompting and evaluation.
Model
GPT-4 / ChatGPT4 / GPT-4 Code Interpreter / GPT-4 Technical Report
Concepts
Failure mode
Methods
pass@k / pass n@k / pass@1 / pass@t [eval]
Related work
HumanEval [builds-on]
Related findings
IC-1414, IC-1415
Extraction
automatic-extraction