SWE-bump-bench is a benchmark for evaluating large language models on real world breaking-change dependency upgrade tasks collected from GitHub. Given a codebase, a package to bump, and package version to bump to, a language model is tasked with generating a patch that resolves any breaking changes for the upgrade.
The dataset contains only repositories consisting primarily of typescript.
First, we collect a list of repos where strict: true is set in a projects tsconfig.json. These repos are added to raw/repos.csv. Next we filter the list to repositories that have at least one major package version to upgrade and which contains breaking changes.
Breaking changes are defined as changes that cause tsc to fail after the package is updated in the repository.
The collect script does this filtering for us:
pnpm run data collect -i raw/repos.csv -o tasks.json
The final output, tasks.json is a JSON file with a collection of valid tasks.
MIT