Can Programming Languages Boost Each Other via Instruction Tuning?

TL;DR

This paper investigates whether instruction fine-tuning with multiple programming languages can mutually enhance code generation performance in large language models, demonstrating significant cross-language improvements.

Contribution

It introduces a comprehensive study showing that training on one programming language can improve performance on others during instruction tuning of large language models.

Findings

Python training improves Java pass@1 by 17.95%

HTML training improves Java pass@1 by 15.24%

Languages can significantly boost each other's code generation capabilities

Abstract

When human programmers have mastered a programming language, it would be easier when they learn a new programming language. In this report, we focus on exploring whether programming languages can boost each other during the instruction fine-tuning phase of code large language models. We conduct extensive experiments of 8 popular programming languages (Python, JavaScript, TypeScript, C, C++, Java, Go, HTML) on StarCoder. Results demonstrate that programming languages can significantly improve each other. For example, CodeM-Python 15B trained on Python is able to increase Java by an absolute 17.95% pass@1 on HumanEval-X. More surprisingly, we found that CodeM-HTML 7B trained on the HTML corpus can improve Java by an absolute 15.24% pass@1. Our training data is released at https://github.com/NL2Code/CodeM.