InfiBench: Evaluating the Question-Answering Capabilities of Code Large Language Models

TL;DR

InfiBench is proposed, the first large-scale freeform question-answering (QA) benchmark for code to the authors' knowledge, comprising 234 carefully selected high-quality Stack Overflow questions that span across 15 programming languages.

Abstract

Large Language Models for code (code LLMs) have witnessed tremendous progress in recent years. With the rapid development of code LLMs, many popular evaluation benchmarks, such as HumanEval, DS-1000, and MBPP, have emerged to measure the performance of code LLMs with a particular focus on code generation tasks. However, they are insufficient to cover the full range of expected capabilities of code LLMs, which span beyond code generation to answering diverse coding-related questions. To fill this gap, we propose InfiBench, the first large-scale freeform question-answering (QA) benchmark for code to our knowledge, comprising 234 carefully selected high-quality Stack Overflow questions that span across 15 programming languages. InfiBench uses four types of model-free automatic metrics to evaluate response correctness where domain experts carefully concretize the criterion for each question. We conduct a systematic evaluation for over 100 latest code LLMs on InfiBench, leading to a series of novel and insightful findings. Our detailed analyses showcase potential directions for further advancement of code LLMs. InfiBench is fully open source at https://infi-coder.github.io/infibench and continuously expanding to foster more scientific and systematic practices for code LLM evaluation.

Figures (5)

• Figure 1: overview. We construct the benchmark by filtering high-quality and diverse question posts from Stack Overflow and annotating question-level evaluation criteria with domain experts. With an model-free automatic evaluation framework, we evaluate over 100 latest code LLMs (one of the most extensive evaluations for code LLMs to the best of our knowledge), leading to several insightful findings.
• Figure 2: A challenging question paraphrased from Stack Overflow where GPT-4 fails to answer.
• Figure 3: Scatter plot of filtered Stack Overflow questions. Questions above the orange line kept.
• Figure 4: Scatter plot for all evaluated LLMs on . $x$-axis is the model size in terms of number of parameters and $y$-axis is score. Projected empirical scaling laws for both general and code models are drawn. Detail discussion in \ref{['subsec:results-and-analysis']}.
• Figure 5: and HumanEval scores as a scatter plot for LLMs. $r=0.8058$. Discussion in \ref{['subapp:correlation-study']}.