Clarification on ChainSWE Gold Replay and F2P/P2P Evaluation Semantics
Hi we have some problem buiding the evaluator as we find in some cases failed when we replay gold patches and test(as stated in paper).1. In the actual ChainSWE evaluation used in the paper, how is a bug stage determined to be resolved?
Is a stage considered resolved as long as the released test_cmds complete successfully with return code 0, or does the evaluator additionally require all declared FAIL_TO_PASS tests to pass and all PASS_TO_PASS tests to remain passing, following the standard SWE-bench grading semantics?
2. If F2P/P2P are part of the actual grading criterion, how does the official evaluator handle declared test IDs that are not executed by the current test_cmds, or cannot be matched to pytest collection / test-log IDs?
Are they treated as failures, ignored, normalized through source-specific mappings, or executed separately?
3. Section 3.2 states that during chain construction and Gold Replay, the test commands are executed and all FAIL_TO_PASS tests are verified to pass while no PASS_TO_PASS tests regress.
How should we interpret released instances where test_cmds appear to cover only a small subset of the declared F2P/P2P tests?
We are not sure whether there is an additional evaluation step that is not directly represented in the released test_cmds field, or whether different source benchmarks use their own test runners or parsers.
4. Finally, is the exact ChainSWE Gold Replay / evaluation runner used for the paper publicly available?
If not, would it be possible to point us to the specific SWE-Edit / SWE-bench evaluator version, as well as the test-log parsing and grading logic used in the experiments? This would help us ensure that our reproduction follows the exact ChainSWE evaluation semantics rather than introducing our own choice between return-code-based and strict F2P/P2P-based grading.
Hi,
Thank you so much for pointing out this issue. We have also received similar feedback from some of our internal authors since the release.
First, to answer your question, we use the standard F2P/P2P evaluation scheme that SWE tasks are expected to follow.
The broader issue is that ChainSWE is built upon six existing SWE-bench datasets, some of which contain noisy instances because their benchmark construction relies heavily on automated agents (see SWE-bench-Live and SWE-rebench). As a result, some instances contain broken or incomplete test environments, or mismatches between test suites and test commands. Some even have different test environments for the same repo (because the docker images are also agent-configured). We need to ensure the basically solvability of these instances to make the on-top chain construction even possible. For your reference, there are similar issues in the initial release of SWE-bench (see issue and issue), which is exactly why SWE-bench was later manually verified and filtered into SWE-bench Verified.
Inspired by this, we are currently working on an updated version of ChainSWE. Instead of relying on unstable and noisy source benchmark environments, the new version expands the size and incorporates manually-validated production environments drawn from our authors’ own development workflows. We expect to release it soon, along with the corresponding evaluation code. Please stay tuned.
Hi,
Thank you very much for the detailed clarification. This is very helpful.
What you described is consistent with what I observed during reproduction, especially the mismatches between the declared F2P/P2P tests, the provided test commands, and the tests that can actually be exercised in some of the released environments. Knowing that part of this comes from noise inherited from the source benchmarks helps clarify the issue a lot.
I’ll revisit my evaluation of the current ChainSWE release with this in mind. I’m also looking forward to the updated version with the manually validated environments and the corresponding evaluation code.
Thanks again for the clarification and for your work on improving the benchmark!
Hi, thanks for the awesome work! We also encountered several issues while reproducing the results. In some instances, the provided patches cannot even be applied. We also noticed that some base commits appear to already include fixes for later tasks in the same chain, potentially exposing their solutions. Additionally, some task descriptions are ambiguous and do not appear to have undergone manual validation.
Besides, the task ordering in some chains appears inconsistent with the repository’s Git history: some later tasks correspond to issues or fixes that actually predate earlier tasks in the chain. This could allow agents to recover solutions from existing commit messages or diffs rather than solve the tasks independently.
Is an updated version of ChainSWE available to use now? If not, is there an estimated release timeline? Thanks!
Hi,
Thanks for the inquiry, we are in the final stage and the updated manually-aligned version will be available by early October.
Best,
ChainSWE Team