SWE-bench
Real GitHub issues that a model must resolve by editing a repository.
Aperçu
SWE-bench takes real issues from open-source repositories and requires a model to produce a patch that makes the project's own tests pass. It solves the problem that isolated programming exercises do not resemble real work, where the task is to find the relevant code in a large repository first. It exists because execution against a real test suite is the hardest-to-game signal for coding ability.
Graphe des technologies
Comment SWE-bench se rattache au reste de la pile.
Appliqué à
- AI Agent application AI Agents SWE-bench used for AI Agent High ✓ source
- Code Generation application Applications SWE-bench used for Code Generation High ✓ source