Contains patches submitted by six repair systems on the swe-bench official website and run logs.

Stores 12 swe-bench-verified code repositories. For demonstration, only the Astropy project is included. In actual use, you should clone the other 11 projects and put them here.

Used to reset the initial state of the projects in swe_projects.

Related to the bug fix experiments mentioned in the paper. Files included:

• draw.py: Run to generate repair_statistics.json, repair_statistics.png, and overall_repair_info.json.

Related to file-level fault localization experiments mentioned in the paper. Files included:

• golden_files.json: Original file.
• get_exp_predictions.py: Run to obtain the sets of error files predicted by different repair methods for each case, output to the exp_predictions folder.
• draw.py: Run to get statistics for hit_at_least_one and hit_all, and generate respective plots.

Related to symbol-level fault localization experiments mentioned in the paper. Files included:

• get_exp_line_info.py: Run to obtain the names of the error files predicted by each repair system.
• get_buggy_files.py: Search and generate the source code of these files.
• code_symbol_parser.py: Perform AST analysis to obtain the actual predicted code symbols.
• draw.py: Used for plotting.

Used to generate issue quality annotations with DeepSeek-r1 and includes five dimensions:

• file_level_gt
• symbol_level_gt
• line_level_gt
• reproduce_gt
• solution_gt

To operate within each folder:

1. Enter the corresponding folder for each dimension and run llm_openrouter.py to get ground truth predictions.
2. Run result_parse.py to obtain the final statistical results.