[In case you are interested in what was fixed in LocAgent and what corner cases remain] Key design choices/data processing code behaviour: 1. We discard all PRs that add or remove files. This is rare in SWE-{Smith, Bench, Gym} (<1% of cases), but quite frequent for SWE-rebench (15% of cases). Our agent cannot predict files that were added by a PR and determining relevant classes and functions is not straightforward for deleted files (by default the code previously detected all functions and classes as relevant) 2. Do not include non-python files in ground truth (our agent is designed to only localize .py files) 3. If PR adds a new function to a class or edits/adds attributes, these changes are detected in ground truth at module and file-level but not at entity-level (previous code used to ignore these changes completely). Function-level docstrings were detected as edits previously --> ignored now. Changes to global variables and imports will only be detected at file-level (as expected). 4. SWE-smith has 11K instances with empty issue description -- these are discarded. Corner cases which we still cannot handle (all these are mostly very rare except for 1): 1. We ignore non-py files from ground truth but it is possible that these are relevant for fixing the issue (for e.g. fixing a .j2 prompt template). (25% of PRs in SWE-Gym and 43% in SWE-rebench edit a non-python file). Should we discard these instances from our dataset? 2. If the ast parser raises an error when reading a file (due to some non-printable characters, syntax errors. etc.), we will discard the instance (rare) 3. If the patch does not apply due to line endings issue (\r\n vs \n) we discard the instance (very rare) 4. There is no way we can detect code refactoring (rearrangement of imports, renaming functions/classes, etc.).