Decaf

About Decaf

Decaf is a research framework for decompilation that utilizes Large Language Models (LLMs) and automatic feedback mechanisms. The system employs LLM generators at various scales (1.3b, 6.7b, and 22b parameters) to sample decompiled code, which is then processed through evaluation and reranking stages. It includes specialized neural rerankers (32b) designed for both stripped and unstripped binaries. The framework provides a pipeline for vulnerability recovery experiments, specifically targeting the Juliet dataset, and supports evaluation against ExeBench test sets. Its core utilities facilitate compiling, executing, disassembling, and calculating edit distance for LLM-generated code. The environment is designed to run within a VS Code Dev Container, utilizing vLLM for inference and requiring HuggingFace authentication for model access.

Summarised from the official page linked above.