MultiPL-E HumanEval (C++)
Code · 2022-08-17
The C++ translation of HumanEval from MultiPL-E, which compiles the Python problems' signatures, docstrings and unit tests into other languages, dropping problems that cannot be translated. A problem is solved when the generated C++ code passes the translated tests; scores are pass@1.
Top models (higher is better)
| Model | Score |
|---|---|
| Llama 3.1 405B Instruct | 82.0 |
| Llama 3.1 70B Instruct | 71.4 |
| Code Llama Python 70B | 54.7 |
| Code Llama Instruct 70B | 53.4 |
| Code Llama 70B | 52.8 |
| Llama 3.1 8B Instruct | 52.8 |
| Code Llama 34B | 47.8 |
| Code Llama Instruct 34B | 45.3 |
| Code Llama Instruct 13B | 42.2 |
| Code Llama Python 34B | 42.2 |
| Code Llama 13B | 39.1 |
| Code Llama Python 13B | 39.1 |
| Code Llama Python 7B | 32.3 |
| Code Llama Instruct 7B | 31.1 |
| Llama 2 70B | 30.4 |
| Code Llama 7B | 28.6 |
| Llama 2 34B | 23.6 |
| Llama 2 13B | 13.7 |
| Llama 2 7B | 6.8 |