investigate DeepSeek-Coder vs. GPT-4o FOR HPC-Fortran-Cpp

[chunhualiao]

Fine-tuned DeekSeek-Coder 6.7B's results look too good when compared to GPT-4 turbo

CodeBLUE is 0.149 < 0.262.

However, the compilation and execution results are the same or better than GPT-4

• compilation: 0.70 == 0.70
• execution: 0.65 > 0.52

When I try to reproduce the results using .cpp files under Fortran2Cpp/data/deepseek_coder_base_6.7B/readable_format_HPEC/cpp

None of them can be compiled. The cpp files contain natural language text and code blocks within ...

https://github.com/bin123apple/Fortran2Cpp/blob/main/data/deepseek_coder_base_6.7B/readable_format_HPEC/cpp/file000.cpp

It looks like these files still need to be processed to extract cpp content.

Where are the extracted cpp files used to generate the results in Table 3? @bin123apple Can you upload them into the git repo's right folder?