harbor run -d ibragim-badertdinov/swe-rebench-v2.1-rlharbor run -d ibragim-badertdinov/swe-rebench-v2.1-rl13,337 verified SWE-rebench V2.1 tasks for RL training of coding agents across 20 languages, with public prebuilt Docker images
harbor run -d ibragim-badertdinov/swe-rebench-v2.1-rl13,337 verified SWE-rebench V2.1 tasks for RL training of coding agents across 20 languages, with public prebuilt Docker images
harbor run -d ibragim-badertdinov/swe-rebench-v2.1-rl13,337 executable software-engineering tasks for RL training of coding agents, in Harbor format. Tasks come from SWE-rebench V2: real GitHub issues with the merged fix as the reference solution and the repository tests as the verifier.
| Tasks | 13,337 |
| Languages | Python (3,990), Go (2,269), TypeScript (1,798), JavaScript (1,687), Rust (1,129), Java (573), Julia (434), PHP (424), Kotlin (306), Elixir (186), Scala (137), Swift (127), Dart (89), R (69), Clojure (53), C# (22), Lua (16), C (15), OCaml (11), C++ (2) |
| Environment | Prebuilt public image per task, pinned by digest |
| Scoring | Binary: reward 1 when all FAIL_TO_PASS tests pass and no PASS_TO_PASS test fails |
tests/test.sh applies tests/test.patch, runs the test command and grades the log with tests/grade.py.tests/grade.py is a standalone standard-library grader; tests/config.json lists FAIL_TO_PASS and PASS_TO_PASS.
Nothing is downloaded at verification time.solution/solve.sh applies the reference patch.[metadata] in task.toml holds repo, language and log_parser.uvx harbor run -d ibragim-badertdinov/swe-rebench-v2.1-rl -a <agent> -m <model>
If you use this dataset, please cite the SWE-rebench V2 paper:
@misc{badertdinov2026swerebenchv2languageagnosticswe,
title={SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale},
author={Ibragim Badertdinov and Maksim Nekrashevich and Anton Shevtsov and Alexander Golubev},
year={2026},
eprint={2602.23866},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2602.23866},
}
13,337 executable software-engineering tasks for RL training of coding agents, in Harbor format. Tasks come from SWE-rebench V2: real GitHub issues with the merged fix as the reference solution and the repository tests as the verifier.
| Tasks | 13,337 |
| Languages | Python (3,990), Go (2,269), TypeScript (1,798), JavaScript (1,687), Rust (1,129), Java (573), Julia (434), PHP (424), Kotlin (306), Elixir (186), Scala (137), Swift (127), Dart (89), R (69), Clojure (53), C# (22), Lua (16), C (15), OCaml (11), C++ (2) |
| Environment | Prebuilt public image per task, pinned by digest |
| Scoring | Binary: reward 1 when all FAIL_TO_PASS tests pass and no PASS_TO_PASS test fails |
tests/test.sh applies tests/test.patch, runs the test command and grades the log with tests/grade.py.tests/grade.py is a standalone standard-library grader; tests/config.json lists FAIL_TO_PASS and PASS_TO_PASS.
Nothing is downloaded at verification time.solution/solve.sh applies the reference patch.[metadata] in task.toml holds repo, language and log_parser.uvx harbor run -d ibragim-badertdinov/swe-rebench-v2.1-rl -a <agent> -m <model>
If you use this dataset, please cite the SWE-rebench V2 paper:
@misc{badertdinov2026swerebenchv2languageagnosticswe,
title={SWE-rebench V2: Language-Agnostic SWE Task Collection at Scale},
author={Ibragim Badertdinov and Maksim Nekrashevich and Anton Shevtsov and Alexander Golubev},
year={2026},
eprint={2602.23866},
archivePrefix={arXiv},
primaryClass={cs.SE},
url={https://arxiv.org/abs/2602.23866},
}