Revue de SWEbench

HumanEval is a benchmark dataset developed by OpenAI specifically for evaluating large language models on code generation tasks, focusing on understanding programming tasks and producing syntactically correct and functionally accurate code.

SWEbench focuses on real-world bug fixes in existing codebases, requiring models to handle long contexts and operate within execution environments. HumanEval, in contrast, primarily assesses the ability to generate standalone functions from docstrings and unit tests, making it a simpler, function-level code generation benchmark.

LiveCodeBench evaluates LLMs on 400 problems from competitive programming platforms, focusing on code generation, self-repair, and test output prediction, with problems updated over time to reduce data contamination.

While SWEbench focuses on fixing real-world bugs in existing repositories, LiveCodeBench emphasizes competitive programming challenges and the ability to self-repair code, often using problems released after a model's training cutoff to ensure genuine generalization.

ClassEval is a manually constructed benchmark that measures how well LLMs can generate full classes of code, including tasks with library, field, or method dependencies, reflecting real-world software engineering scenarios.

SWEbench evaluates bug-fixing capabilities within large, existing codebases, whereas ClassEval specifically assesses the generation of complete, interdependent code classes, moving beyond isolated functions to more complex structural coding tasks.

APPS is a large-scale code generation benchmark comprising 10,000 problems collected from open-access competitive coding websites, ranging from one-line solutions to substantial algorithmic challenges.

SWEbench is centered on resolving real-world software issues and generating patches for bugs in existing repositories. APPS, conversely, evaluates an LLM's ability to generate satisfactory Python code from natural language specifications, primarily focusing on algorithmic problem-solving rather than bug fixing in a pre-existing codebase.

This benchmark evaluates LLMs on real-world software engineering tasks sourced directly from Upwork freelance jobs, including both coding ability and engineering management decisions, with actual dollar values attached.

Both SWEbench and this benchmark focus on real-world software engineering problems. However, the Upwork benchmark uniquely ties performance to economic value and includes higher-level engineering management decisions, whereas SWEbench is specifically focused on generating patches to fix GitHub issues.