DevEval: A Manually-Annotated Code Generation Benchmark Aligned with Real-World Code Repositories

How to evaluate the coding abilities of Large Language Models (LLMs) remains an open question. We find that existing benchmarks are poorly aligned with real-world code repositories and are insufficient to evaluate the coding abilities of LLMs. To address the knowledge gap, we propose a new benchmark named DevEval, which has three advances. (1) DevEval aligns with real-world repositories in multiple dimensions, e.g., code distributions and dependency distributions. (2) DevEval is annotated by 13 developers and contains comprehensive annotations (e.g., requirements, original repositories, reference code, and reference dependencies). (3) DevEval comprises 1,874 testing samples from 117 repositories, covering 10 popular domains (e.g., Internet, Database). Based on DevEval, we propose repository-level code generation and evaluate 8 popular LLMs on DevEval (e.g., gpt-4, gpt-3.5, StarCoder 2, DeepSeek Coder, CodeLLaMa). Our experiments reveal these LLMs' coding abilities in real-world code repositories. For example, in our experiments, the highest Pass@1 of gpt-4-turbo is only 53.04%. We also analyze LLMs' failed cases and summarize their shortcomings. We hope DevEval can facilitate the development of LLMs in real code repositories. DevEval, prompts, and LLMs' predictions have been released.

DevEval: Evaluating Code Generation in Practical Software Projects

How to evaluate Large Language Models (LLMs) in code generation is an open question. Many benchmarks have been proposed but are inconsistent with practical software projects, e.g., unreal program distributions, insufficient dependencies, and small-scale project contexts. Thus, the capabilities of LLMs in practical projects are still unclear. In this paper, we propose a new benchmark named DevEval, aligned with Developers' experiences in practical projects. DevEval is collected through a rigorous pipeline, containing 2,690 samples from 119 practical projects and covering 10 domains. Compared to previous benchmarks, DevEval aligns to practical projects in multiple dimensions, e.g., real program distributions, sufficient dependencies, and enough-scale project contexts. We assess five popular LLMs on DevEval (e.g., gpt-4, gpt-3.5-turbo, CodeLLaMa, and StarCoder) and reveal their actual abilities in code generation. For instance, the highest Pass@1 of gpt-3.5-turbo only is 42 in our experiments. We also discuss the challenges and future directions of code generation in practical projects. We open-source DevEval and hope it can facilitate the development of code generation in practical projects.

Towards Enhancing In-Context Learning for Code Generation

In-context learning (ICL) with pre-trained language models (PTLMs) has shown great success in code generation. ICL does not require training. PTLMs take as the input a prompt consisting of a few requirement-code examples and a new requirement, and output a new program. However, existing studies simply reuse ICL techniques for natural language generation and ignore unique features of code generation. We refer to these studies as standard ICL. Inspired by observations of the human coding process, we propose a novel ICL approach for code generation named AceCoder. Compared to standard ICL, AceCoder has two novelties. (1) Example retrieval. It retrieves similar programs as examples and learns programming skills (e.g., algorithms, APIs) from them. (2) Guided Code Generation. It encourages PTLMs to output an intermediate preliminary (e.g., test cases, APIs) before generating programs. The preliminary can help PTLMs understand requirements and guide the next code generation. We apply AceCoder to six PTLMs (e.g., Codex) and evaluate it on three public benchmarks using the Pass@k. Results show that AceCoder can significantly improve the performance of PTLMs on code generation. (1) In terms of Pass@1, AceCoder outperforms standard ICL by up to 79.7% and fine-tuned models by up to 171%. (2) AceCoder is effective in PTLMs with different sizes (e.g., 1B to 175B) and different languages (e.g., Python, Java, and JavaScript). (3) We investigate multiple choices of the intermediate preliminary. (4) We manually evaluate generated programs in three aspects and prove the superiority of AceCoder. (5) Finally, we discuss some insights about ICL for practitioners.