README.md

1. Introduction of Falcon3-decompile-3b

Falcon3-decompiler-3b aims to decompile x86 assembly instructions into C.

2. Evaluation Results

3. How to Use

Here is an example of how to use our model Note: Replace asm_func with the function that you want to decompile

Decompilation: Use falcon3-decompiler-3b to translate ghidra decompilation output to more readable code:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model_path = 'LLM4Binary/llm4decompile-1.3b-v1.5' # V1.5 Model
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path,torch_dtype=torch.bfloat16).cuda()

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
import os

asm_func = """
char * func0(char **param_1,int param_2)

{
  char **ppcVar1;
  char *__s;
  size_t sVar2;
  int iVar3;
  char *pcVar4;
  
  pcVar4 = "";
  if (0 < param_2) {
    iVar3 = 0;
    ppcVar1 = param_1 + (ulong)(param_2 - 1) + 1;
    do {
      __s = *param_1;
      sVar2 = strlen(__s);
      if (iVar3 < (int)sVar2) {
        pcVar4 = __s;
        iVar3 = (int)sVar2;
      }
      param_1 = param_1 + 1;
    } while (param_1 != ppcVar1);
  }
  return pcVar4;
}
"""

before = f"# This is the assembly code:\n"#prompt
after = "\n# What is the source code?\n"#prompt
asm_func = before+asm_func.strip()+after
model_path = "Neo111x/falcon3-decompiler-3b"
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForCausalLM.from_pretrained(model_path, torch_dtype="auto", device_map="auto").to("cuda:0")

inputs = tokenizer(asm_func, return_tensors="pt").to("cuda:0")
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])

# Note only decompile one function, where the original file may contain multiple functions

print(f'decompiled function:\n{c_func_decompile}')

4. Contact

If you have any questions, please raise an issue.