BlazingCustoms/pybytecode-v3-1.5b
PyBytecode v3 is a 1.5 billion parameter model by BlazingCustoms, fine-tuned from Qwen2.5-Coder-1.5B-Instruct, specifically designed to decompile Python 3.12 bytecode back into source code. This model offers a unique verification mechanism, allowing users to recompile its output and compare the bytecode for exact matches, ensuring high confidence in the decompiled source. It excels at accurately reconstructing Python source from bytecode, particularly for individual functions under 100 disassembly lines.
Loading preview...
Overview
PyBytecode v3 is a specialized 1.5 billion parameter language model developed by BlazingCustoms, fine-tuned from Qwen2.5-Coder-1.5B-Instruct. Its core function is to accurately decompile Python 3.12 bytecode into human-readable Python source code. A key differentiator is its built-in verification method, enabling users to compile the model's output and compare the resulting bytecode against the original, providing a definitive "certified" answer.
Key Capabilities
- Python 3.12 Bytecode Decompilation: Converts disassembled Python 3.12 code objects back into their original source.
- Verifiable Output: Allows for recompilation and bytecode comparison to confirm the exact correctness of the decompiled source, achieving up to 93.67% certification rate with multiple attempts.
- Performance: Achieves 92.86% certified accuracy on functions with under 100 disassembly lines in a single attempt, and 98.21% with up to 32 attempts.
- Apache-2.0 Licensed: Weights are openly available under an Apache-2.0 license, with GGUF builds provided for broader compatibility.
Limitations and Considerations
- Input Length Sensitivity: Performance significantly degrades with longer inputs; it is not effective for functions exceeding ~400 disassembly lines.
- Python 3.12 Specific: Trained and tested exclusively on Python 3.12 bytecode.
- Compilation Flags: Requires careful handling of Python's
-Ocompilation flag to ensure accurate verification. - Not for Malware Analysis: Has not been shown to work on real-world packed or obfuscated malware.