Generating Unseen Code Tests In Infinitum

TL;DR

This paper introduces a method for generating diverse, unseen code tests to evaluate large language models' coding abilities, addressing data leakage issues in benchmarks and supporting ongoing model assessment.

Contribution

The authors propose a novel approach for creating adaptable, non-leaking benchmark variations applicable across coding tasks and languages, including in-house code bases.

Findings

Developed 'auto-regression' benchmark for Python code generation.

Enables continuous testing and debugging of LLMs.

Mitigates data leakage in model evaluation.

Abstract

Large Language Models (LLMs) are used for many tasks, including those related to coding. An important aspect of being able to utilize LLMs is the ability to assess their fitness for specific usages. The common practice is to evaluate LLMs against a set of benchmarks. While benchmarks provide a sound foundation for evaluation and comparison of alternatives, they suffer from the well-known weakness of leaking into the training data \cite{Xu2024Benchmarking}. We present a method for creating benchmark variations that generalize across coding tasks and programming languages, and may also be applied to in-house code bases. Our approach enables ongoing generation of test-data thus mitigating the leaking into the training data issue. We implement one benchmark, called \textit{auto-regression}, for the task of text-to-code generation in Python. Auto-regression is specifically created to aid in debugging and in tracking model generation changes as part of the LLM regression testing process.