LFM2.5-DCR-8B-A1B-GGUF
This model is a GGUF quantized version of
LFM2.5-DCR-8B-A1B, which is a fine-tune of
LFM 2.5, specialized to be a
DCR, or
Decompiled Code Reconstructor, for the resulting C++ code produced by Unity's IL2CPP build process.
I picked the 8B version because it is a Mixture-of-Experts model with only about 1 billion parameters active at a time, so reconstruction can be fast while still using one of the most impressive models in its parameter range.
It is trained on 200+ pieces of decompiled Unity IL2CPP snippets and the correct corresponding C# output. Its goal is to reconstruct clean, readable C# from IL2CPP-style decompiled C++ pseudocode.
This is the quantized 4-bit
Q4_K_M GGUF version. You can find the full precision 16-bit version
here.
Intended Use
This model is intended for reconstructing C# source-like code from decompiled Unity IL2CPP C++ output.
Example use cases:
- Reconstructing Unity MonoBehaviour methods from IL2CPP pseudocode
- Translating decompiler-style C++ control flow back into readable C#
- Cleaning up IL2CPP field access, method calls, and Unity API patterns
- Assisting with code understanding and reverse-engineering workflows where permitted
Training Data
The full precision model was fine-tuned on 200+ curated examples of decompiled Unity IL2CPP snippets paired with correct corresponding C# reconstructions.
Training examples used a reasoning-plus-answer format, where the model learns to analyze the decompiled IL2CPP structure and then produce clean final C# code.
Base Models
Fine-tuned full precision model:
famseamus/LFM2.5-DCR-8B-A1BOriginal upstream model:
LiquidAI/LFM2.5-8B-A1BOriginal upstream GGUF reference:
LiquidAI/LFM2.5-8B-A1B-GGUFVersion
This repository contains the 4-bit Q4_K_M GGUF quantized version for local inference with GGUF-compatible runtimes such as llama.cpp.
For full precision use, see:
