albertan017
Where is the decompile-eval-executable-gcc-ghidra.json
Thank you for your fantastic work! we noticed that in your colab scripts there is a /LLM4Decompile/decompile-eval/decompile-eval-executable-gcc-ghidra.json file. But we didnot find this file in the project, can you provide it?
I find the file in the LLM4Decompile/legacy-test/! by the way, we found that in the LLM4Decompile-Ref part of the colab script, the prompt is before = "# This is the assembly code:\n" after = "\n# What is the source code?\n". We wonder why isn't the prompt "# This is the pseudocode:\n" since LLM4Decompile-Ref takes the ghidra'pseudocode as input
Thank you for the question. We marked this as a legacy test because we have integrated the dataset into decompile-bench (https://huggingface.co/datasets/LLM4Binary/decompile-eval/tree/main), which now includes the HumanEval, MBPP, and Github2025 datasets.
Regarding the prompt, that is a legacy artifact. However, since we used this exact prompt structure during model training, the specific wording is less critical than the consistency between training and inference
