LLM4Decompile

Reverse Engineering: Decompiling Binary Code with Large Language Models

Updates

• [2025-10-04]: Release SK²Decompile: LLM-based Two-Phase Binary Decompilation from Skeleton to Skin. Phase 1 Structure Recovery (Skeleton): Transform binary/pseudo-code into obfuscated intermediate representations 🤗 HF Link. Phase 2 Identifier Naming (Skin): Generate human-readable source code with meaningful identifiers 🤗 HF Link.
• [2025-05-20]: Release decompile-bench, contains two million binary-source function pairs for training, and 70K function pairs for evaluation. Please refer to the decompile-bench folder for details.
• [2024-10-17]: Release decompile-ghidra-100k, a subset of 100k training samples (25k per optimization level). We provide a training script that runs in ~3.5 hours on a single A100 40G GPU. It achieves a 0.26 re-executability rate, with a total cost of under $20 for quick replication of LLM4Decompile.
• [2024-09-26]: Update a Colab notebook to demonstrate the usage of the LLM4Decompile model, including examples for the LLM4Decompile-End and LLM4Decompile-Ref models.
• [2024-09-23]: Release LLM4Decompile-9B-v2, fine-tuned based on Yi-Coder-9B, achieved a re-executability rate of 0.6494 on the Decompile benchmark.
• [2024-06-19]: Release V2 series (LLM4Decompile-Ref). V2 (1.3B-22B), building upon Ghidra, are trained on 2 billion tokens to refine the decompiled pseudo-code from Ghidra. The 22B-V2 version outperforms the 6.7B-V1.5 by an additional 40.1%. Please check the ghidra folder for details.
• [2024-05-13]: Release V1.5 series (LLM4Decompile-End, directly decompile binary using LLM). V1.5 are trained with a larger dataset (15B tokens) and a maximum token length of 4,096, with remarkable performance (over 100% improvement) compared to the previous model.
• [2024-03-16]: Add llm4decompile-6.7b-uo model which is trained without prior knowledge of the optimization levels (O0~O3), the average re-executability is around 0.219, performs the best in our models.

About

• LLM4Decompile is the pioneering open-source large language model dedicated to decompilation. Its current version supports decompiling Linux x86_64 binaries, ranging from GCC's O0 to O3 optimization levels, into human-readable C source code. Our team is committed to expanding this tool's capabilities, with ongoing efforts to incorporate a broader range of architectures and configurations.
• LLM4Decompile-End focuses on decompiling the binary directly. LLM4Decompile-Ref refines the pseudo-code decompiled by Ghidra.

Evaluation

Framework

During compilation, the Preprocessor processes the source code (SRC) to eliminate comments and expand macros or includes. The cleaned code is then forwarded to the Compiler, which converts it into assembly code (ASM). This ASM is transformed into binary code (0s and 1s) by the Assembler. The Linker finalizes the process by linking function calls to create an executable file. Decompilation, on the other hand, involves converting binary code back into a source file. LLMs, being trained on text, lack the ability to process binary data directly. Therefore, binaries must be disassembled by Objdump into assembly language (ASM) first. It should be noted that binary and disassembled ASM are equivalent, they can be interconverted, and thus we refer to them interchangeably. Finally, the loss is computed between the decompiled code and source code to guide the training. To assess the quality of the decompiled code (SRC'), it is tested for its functionality through test assertions (re-executability).

Metrics

• Re-executability evaluates whether the decompiled code can execute properly and pass all the predefined test cases.

Benchmarks

• HumanEval-Decompile A collection of 164 C functions that exclusively rely on standard C libraries.
• ExeBench A collection of 2,621 functions drawn from real projects, each utilizing user-defined functions, structures, and macros.

Results

Models

Our LLM4Decompile includes models with sizes between 1.3 billion and 33 billion parameters, and we have made these models available on Hugging Face.

Model	Checkpoint	Size	Re-executability	Note
llm4decompile-1.3b-v1.5	🤗 HF Link	1.3B	27.3%	Note 3
llm4decompile-6.7b-v1.5	🤗 HF Link	6.7B	45.4%	Note 3
llm4decompile-1.3b-v2	🤗 HF Link	1.3B	46.0%	Note 4
llm4decompile-6.7b-v2	🤗 HF Link	6.7B	52.7%	Note 4
llm4decompile-9b-v2	🤗 HF Link	9B	64.9%	Note 4
llm4decompile-22b-v2	🤗 HF Link	22B	63.6%	Note 4

Note 3: V1.5 series are trained with a larger dataset (15B tokens) and a maximum token size of 4,096, with remarkable performance (over 100% improvement) compared to the previous model.

Note 4: V2 series are built upon Ghidra and trained on 2 billion tokens to refine the decompiled pseudo-code from Ghidra. Check ghidra folder for details.

Quick Start

Setup: Please use the script below to install the necessary environment.

git clone https://github.com/albertan017/LLM4Decompile.git
cd LLM4Decompile
conda create -n 'llm4decompile' python=3.9 -y
conda activate llm4decompile
pip install -r requirements.txt

Here is an example of how to use our model (Revised for V1.5. For previous models, please check the corresponding model page at HF). Note: Replace the "func0" with the function name you want to decompile.

Preprocessing: Compile the C code into binary, and disassemble the binary into assembly instructions.

import subprocess
import os
func_name = 'func0'
OPT = ["O0", "O1", "O2", "O3"]
fileName = 'samples/sample' #'path/to/file'
for opt_state in OPT:
    output_file = fileName +'_' + opt_state
    input_file = fileName+'.c'
    compile_command = f'gcc -o {output_file}.o {input_file} -{opt_state} -lm'#compile the code with GCC on Linux
    subprocess.run(compile_command, shell=True, check=True)
    compile_command = f'objdump -d {output_file}.o > {output_file}.s'#disassemble the binary file into assembly instructions
    subprocess.run(compile_command, shell=True, check=True)
    
    input_asm = ''
    with open(output_file+'.s') as f:#asm file
        asm= f.read()
        if '<'+func_name+'>:' not in asm: #IMPORTANT replace func0 with the function name
            raise ValueError("compile fails")
        asm = '<'+func_name+'>:' + asm.split('<'+func_name+'>:')[-1].split('\n\n')[0] #IMPORTANT replace func0 with the function name
        asm_clean = ""
        asm_sp = asm.split("\n")
        for tmp in asm_sp:
            if len(tmp.split("\t"))<3 and '00' in tmp:
                continue
            idx = min(
                len(tmp.split("\t")) - 1, 2
            )
            tmp_asm = "\t".join(tmp.split("\t")[idx:])  # remove the binary code
            tmp_asm = tmp_asm.split("#")[0].strip()  # remove the comments
            asm_clean += tmp_asm + "\n"
    input_asm = asm_clean.strip()
    before = f"# This is the assembly code:\n"#prompt
    after = "\n# What is the source code?\n"#prompt
    input_asm_prompt = before+input_asm.strip()+after
    with open(fileName +'_' + opt_state +'.asm','w',encoding='utf-8') as f:
        f.write(input_asm_prompt)

Assembly instructions should be in the format:

<FUNCTION_NAME>:\nOPERATIONS\nOPERATIONS\n

Typical assembly instructions may look like this:

<func0>:
endbr64
lea    (%rdi,%rsi,1),%eax
retq

Decompilation: Use LLM4Decompile to translate the assembly instructions into C:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_path = 'LLM4Binary/llm4decompile-6.7b-v1.5' # V1.5 Model
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path,torch_dtype=torch.bfloat16).cuda()

with open(fileName +'_' + OPT[0] +'.asm','r') as f:#optimization level O0
    asm_func = f.read()
inputs = tokenizer(asm_func, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=2048)### max length to 4096, max new tokens should be below the range
c_func_decompile = tokenizer.decode(outputs[0][len(inputs[0]):-1])

with open(fileName +'.c','r') as f:#original file
    func = f.read()

print(f'original function:\n{func}')# Note we only decompile one function, where the original file may contain multiple functions
print(f'decompiled function:\n{c_func_decompile}')

Docker setup

# build docker
docker build -t llm4decompile .

# run docker with GPU
docker run --gpus all -it --name llm4decompile llm4decompile /bin/bash

# run demo.py (choose a model suitable for your resources before running)
cd ghidra
python demo.py

HumanEval-Decompile

Data are stored in llm4decompile/decompile-eval/decompile-eval-executable-gcc-obj.json, using JSON list format. There are 164*4 (O0, O1, O2, O3) samples, each with five keys:

• task_id: indicates the ID of the problem.
• type: the optimization stage, is one of [O0, O1, O2, O3].
• c_func: C solution for HumanEval problem.
• c_test: C test assertions.
• input_asm_prompt: assembly instructions with prompts, can be derived as in our preprocessing example.

Please check the evaluation scripts.

On Going

• Larger training dataset with the cleaning process. (done:2024.05.13)
• Support for popular languages/platforms and settings.
• Support for executable binaries. (done:2024.05.13)
• Integration with decompilation tools (e.g., Ghidra, Rizin)

License

This code repository is licensed under the MIT and DeepSeek License.

Citation

@misc{tan2024llm4decompile,
      title={LLM4Decompile: Decompiling Binary Code with Large Language Models}, 
      author={Hanzhuo Tan and Qi Luo and Jing Li and Yuqun Zhang},
      year={2024},
      eprint={2403.05286},
      archivePrefix={arXiv},
      primaryClass={cs.PL}
}

Star History