n64dllm-v1

A LoRA fine-tune of Qwen3-Coder-30B-A3B-Instruct for N64 matching decompilation: given MIPS assembly, write C that recompiles byte-identically under the original SGI IDO 5.3/7.1 compilers. Trained on roughly 74k examples mined from 11 community decomp projects. Every training pair was verified by recompiling it with the original compiler and byte-comparing against the shipped ROM, and every number below is judged the same way. No similarity metrics anywhere.

Results

Held-out games the model never trained on (Pokemon Snap, Mischief Makers), identical prompts for all models, byte-exact recompilation as the judge:

Model	exact match @1	pass@8
Qwen3-Coder-30B-A3B, stock	6.7%
n64dllm-v1, this Q4_K_M file	18.3%	20.0%
Claude Opus 4.8	20.0%

The MoE base has 3.3B active parameters, so this runs at about 90 tok/s on an M4 Max. Cheap sampling plus a compiler oracle is the intended usage pattern.

Files

• n64dllm-v1-Q4_K_M.gguf: for llama.cpp, LM Studio, or any GGUF runtime. Quantization was gated by re-running the benchmark on this exact file; it lost nothing vs bf16.
• adapter/: the LoRA (rank 64, attention projections) if you want to keep training.

Usage

llama-server -m n64dllm-v1-Q4_K_M.gguf --port 8081 -ngl 99 -c 16384 --jinja -fa on

The prompt format matters. Supply the declared prototype from the project's headers; in ablation it was worth 11.7 points:

You are an expert N64 matching-decompilation model. Given MIPS assembly produced by the ido5.3 compiler, output C that recompiles to a byte-identical match.

; compiler: ido5.3
; declared prototype: void func_80123456(Actor* this, s32 arg1);
; target assembly:
<output of objdump -dr --disassemble=<fn> on the target object>

; matching C:

You will get the best results inside an agent harness with the compiler in the loop: sample several candidates, recompile each, byte-compare, feed diffs back. The project repo ships a ready-made skill file for Codex and OpenCode plus a verification harness built on n64-decomp-workbench.

Limitations

Function signatures should be supplied, not guessed; cross-game struct internals are the main remaining failure mode. Assembly beyond ~4k tokens was not trained on. v1 covers IDO games only, not the GCC-family titles. General coding ability of the base survives the fine-tune (the LoRA touches attention projections only), and tool calling works.

Intended use

Research and community decompilation assistance. Matching decompilation is an existing community practice aimed at interoperability and preservation; this model writes new C and contains no game assets. Not for commercial use of decompiled output. The training data is not distributed.