---

name: llm4decompile description: "Route requests for LLM4Decompile training, direct decompilation, Ghidra refinement, and SK²Decompile workflows." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

LLM4Decompile

LLM4Decompile is a decompilation repo with four major user-facing workflow families: training, direct inference/evaluation, Ghidra-assisted refinement, and the SK²Decompile two-phase pipeline. Use this root skill as the router, then open the sub-skill that matches the task shape.

Quick route map

• Fine-tune or continue pretraining models → sub-skills/training/SKILL.md
• Run assembly-to-C inference or benchmark scoring → sub-skills/evaluation/SKILL.md
• Use Ghidra to extract pseudo-code and refine it with a model → sub-skills/ghidra-refine/SKILL.md
• Run the SK²Decompile skeleton/skin pipeline, RL helpers, or BringUpBench evaluation → sub-skills/sk2decompile/SKILL.md

Common signals

Use the workflow that matches the user's nouns and commands:

• train, fine-tune, pretrain, DeepSpeed, ColossalAI, LLaMA-Factory
• vLLM, text-generation-inference, single GPU, HumanEval-Decompile, Decompile-Bench, ExeBench
• Ghidra, analyzeHeadless, pseudo-code, V2, refine
• SK²Decompile, skeleton, skin, GRPO, reward function, BringUpBench

Shared repository facts

• The repo is source-only; it does not declare a Python package in pyproject.toml or setup.py.
• GPU/CUDA is required for the model-training and model-inference paths.
• Java 17 and clang-format are required for the Ghidra and normalization workflows.
• GCC/binutils are required for dataset compilation and benchmark execution.
• Psychec/stack appears only in the SK²Decompile header-inference path.

Before editing or extending the skill tree, read:

• references/model-overview.md
• references/troubleshooting.md
• references/repo-provenance.md
• references/repo-routing-metadata.json
• scripts/check-runtime.sh

How to use this skill

1. Identify the workflow family.
2. Open the matching sub-skill.
3. Use the bundled references for command templates, data schemas, and troubleshooting.
4. Prefer the bundled scripts over source-checkout paths when giving future instructions.

What not to do

• Do not route training questions into the benchmark or Ghidra routes.
• Do not route benchmark questions into the training route.
• Do not route SK²Decompile questions into the direct-evaluation route just because both use CUDA.
• Do not assume the old legacy benchmark examples are the primary path; treat them as historical context only.

Shared helper

Run scripts/check-runtime.sh when you need a quick environment sanity check for the common Python, CUDA, compiler, and Java prerequisites.