README.md

metadata

title: N64LLMDecompile
emoji: 🎮
colorFrom: indigo
colorTo: purple
sdk: gradio
app_file: app.py
pinned: false
license: apache-2.0
tags:
  - build-small-hackathon
  - backyard-ai
  - thousand-token-wood
  - best-demo
  - best-use-of-modal
  - best-use-of-codex
  - track:backyard
  - track:wood
  - sponsor:openai
  - sponsor:nvidia
  - sponsor:modal
  - achievement:offgrid
  - achievement:welltuned
  - achievement:fieldnotes

N64LLMDecompile 🎮

Decompile Nintendo 64 MIPS functions back into C with a fine-tuned LLM, with compiler-verified scoring against the original target objects.

Built for the Hugging Face Build Small hackathon.

Links:

• 💻 Source code: github.com/MatthewLReingold/N64LLMDecompile
• 🎥 Demo video: YouTube link
• 📝 Writeup: blog link
• 📣 Social post: LinkedIn Post

What it does

Given the Ghidra pseudo-C of an N64 MIPS function, the model rewrites it into C source. Optionally, the Space compiles that C with the original compiler and flags and diffs the result against the target object using asm-differ — giving a byte-level match score, where 0 means a byte-perfect match (the metric the decomp community uses to confirm a function is "matched").

Two ways to use it:

• Paste Ghidra pseudo-C → get C back.
• Also upload the target object and pick the compiler + flags → get a match score.

The model

MatthewReingold/llm4decompile-9b-v2-n64-finetune — a fine-tune of LLM4Decompile-9B-v2 on verified N64 scratches from decomp.me.

Results & writeup

Full results, methodology, and what I learned are covered in the blog post:

[ writeup link coming soon ]

Acknowledgments

Thank you to the work of these teams who all gave the resources used during development:

• decomp.me — the scratch database and compiler-toolchain packaging the dataset is built from
• decompals — IDO static recompilation, and the MIPS GCC / binutils builds
• Paper Mario team (pmret) — the GCC 2.8.1 compiler
• LLM4Decompile (LLM4Binary) — the base model this fine-tune is built on
• asm-differ (simonlindholm) — the assembly diffing used for scoring
• Ghidra (NSA) — pseudo-C generation in the data pipeline
• NVIDIA — the Nemotron models, used during experimentation
• Modal — cloud GPU infrastructure and compute credits
• OpenAI — Codex, used during development
• Hugging Face — hosting the hackathon and ZeroGPU access

License

Apache-2.0. The base model and the decomp toolchains retain their own respective licenses.