2026-10-08T21:22:53Z https://api.figshare.com/v2/oai

oai:figshare.com:article/34068783 2026-10-05T09:26:45Z category_29200 category_29203 item_type_3 month_year_10_2026

Replication package for "When Does LLM Test Repair Destroy Bug-Revealing Tests? A Cost-Accounted Study on Real Repository Defects" Zehao Li (25144737) Software testing, verification and validation Software engineering not elsewhere classified software testing test repair test oracle large language models bug reproduction SWT-bench empirical study <p dir="ltr">Replication package for an empirical study of how iterative LLM test repair destroys verified bug-revealing (fail-to-pass) tests on real repository defects (SWT-bench Verified, 229 instances, three LLMs). It contains the study protocol with its dated amendments, the Docker-free execution harness, every LLM request and response with provider-reported cost, every recorded test version with its outcomes on the buggy and the fixed program version, the test-execution logs, the model-based audit samples, the official-harness cross-check, and the scripts that regenerate every table, figure and in-text number of the paper. Benchmark metadata are not redistributed: src/fetch_data.py downloads them from Hugging Face, and work/data_checksums.json verifies them. Code: MIT. Data: CC BY 4.0.</p> 2026-10-05T09:26:45Z Dataset Dataset 10.6084/m9.figshare.34068783.v1 https://figshare.com/articles/dataset/Replication_package_for_When_Does_LLM_Test_Repair_Destroy_Bug-Revealing_Tests_A_Cost-Accounted_Study_on_Real_Repository_Defects_/34068783 CC BY 4.0