Prompting of different code languages?
Taking quick_sort as an example, what is the prompting of different code languages during the inference.
For python, starcoder can directly generate expected output using # quick sort. But, it went wrong when using other languages. How we can recognize the requirements from different code languages.
Is the prompting like // language: c++\n and # language: Python\n. In this condition, the test results is satisfied.
We did condition on filename during pretraining so you can try appending: <filename>file_path.ext\n where ext is the extension of the language you want to generate the code in, you can change the filepath as you want.
For example we found <filename>solutions/solution_1.py\n# Here is the correct implementation of the code exercise\n) to help with solving HumanEval problems in Python.
We did condition on filename during pretraining so you can try appending:
<filename>file_path.ext\nwhere ext is the extension of the language you want to generate the code in, you can change the filepath as you want.
For example we found<filename>solutions/solution_1.py\n# Here is the correct implementation of the code exercise\n) to help with solving HumanEval problems in Python.
Thank you! But in FIM mode, should I add that before or after <fim_prefix>?
Like
<filename>solutions/solution_1.py
<fim_prefix>...<fim_suffix>...
or
<fim_prefix><filename>solutions/solution_1.py
...<fim_suffix>...