CodableLLM: Automating Decompiled and Source Code Mapping for LLM Dataset Generation

TL;DR

CodableLLM is a Python framework that automates the process of mapping decompiled binaries to source code, enabling the creation of high-quality datasets for training large language models in code understanding.

Contribution

It introduces a novel automated method for aligning decompiled and source code, supporting multiple languages and improving dataset quality for code-focused LLMs.

Findings

Efficient dataset generation for code understanding models

Improved alignment accuracy between decompiled and source code

Outperforms existing tools in dataset creation efficiency

Abstract

The generation of large, high-quality datasets for code understanding and generation remains a significant challenge, particularly when aligning decompiled binaries with their original source code. To address this, we present CodableLLM, a Python framework designed to automate the creation and curation of datasets by mapping decompiled functions to their corresponding source functions. This process enhances the alignment between decompiled and source code representations, facilitating the development of large language models (LLMs) capable of understanding and generating code across multiple abstraction levels. CodableLLM supports multiple programming languages and integrates with existing decompilers and parsers to streamline dataset generation. This paper presents the design and implementation of CodableLLM, evaluates its performance in dataset creation, and compares it to existing tools in the field. The results demonstrate that CodableLLM offers a robust and efficient solution for generating datasets tailored for code-focused LLMS.