SWE-Gym

SWE-Gym is a training environment of real Python issue-resolution tasks; agents trained on it are typically reported by resolve rate on SWE-bench Verified and Lite, not a separate SWE-Gym leaderboard.

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categorycoding
Subcategoryagent training environment with held-out evaluation
Page statusactive
Metric% resolved (patch passes FAIL_TO_PASS and PASS_TO_PASS tests)
Directionhigher_is_better
Unit%
Dataset size2438
Dataset licenceApache-2.0

What it measures

SWE-Gym provides real-world Python task instances (a codebase snapshot, an executable runtime, unit tests and a natural-language issue) for training and verifying software-engineering agents. It is built to be run against, not just read: an agent explores the repository, edits code and executes tests inside the provided environment. The paper's own reported numbers measure what training on SWE-Gym does for resolve rate on the separate SWE-bench Verified and Lite test sets, plus a "Lite" 234-instance subset of SWE-Gym itself used for lighter-weight evaluation during development.

Task format

Text input (issue description) with repository and executable test access; output is a patch, graded by running hidden tests recovered from the originating pull request.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub