Divij/Qwen3.5-4B-SFT-decompiler-O0-full
Divij/Qwen3.5-4B-SFT-decompiler-O0-full is a 4.5 billion parameter Qwen3.5-4B model fine-tuned for binary decompilation, specifically to recover original C source code from decompiler output. It excels at transforming generic, obfuscated decompiled C code into clean, idiomatic C with meaningful variable names and proper types. This model is optimized for code compiled with no optimization (O0) and is highly effective for security research, reverse engineering, and binary analysis tasks.
Loading preview...
Model Overview
Divij/Qwen3.5-4B-SFT-decompiler-O0-full is a specialized 4.5 billion parameter language model based on the Qwen3.5-4B architecture. It has been fine-tuned to address the challenging task of binary decompilation, specifically to reconstruct human-readable C source code from the output of various decompilers.
Key Capabilities
- Source Code Recovery: Transforms raw decompiled C code (often with generic variable names, hex constants, and flattened control flow) into clean, idiomatic C code with proper types and meaningful identifiers.
- Optimization Level Focus: Specifically trained on code compiled with no optimization (O0), making it highly effective for binaries compiled under these conditions.
- Multi-Decompiler Support: Trained using outputs from popular decompilers including Ghidra, IDA Pro, and angr (dream, phoenix, sailr variants).
- High Structural Accuracy: Achieves a 75% perfect Control Flow Graph (CFG) match (GED=0) with the original source code, significantly outperforming raw decompiler outputs (34.9%).
- Robust Output: Demonstrates a 99.9% success rate in producing valid, parseable C code, crucial for downstream analysis.
When to Use This Model
This model is ideal for tasks requiring the reconstruction of high-quality C source code from binaries, particularly in:
- Security Research: Analyzing malware or proprietary software to understand its functionality.
- Reverse Engineering: Gaining insights into compiled applications where source code is unavailable.
- Binary Analysis: Improving the readability and maintainability of decompiler outputs for further manual or automated analysis.
It offers substantial improvements over raw decompiler outputs, especially in structural fidelity and code readability, making it a valuable tool for anyone working with compiled C programs.