bookxd/jmh-grpo-replication-data

JMH-Bench GRPO replication data

Data for replicating the BSc thesis Reinforcement Learning with Verifiable Execution Rewards for Synthesizing Regression-Detecting Microbenchmarks. The code is in the replication repository RL4JMH-thesis-submission. The trained models are run 9 (evaluated in the thesis) and run 8.

Path	Contents
classpaths/	Mutant-bearing jars and dependency jars of the 246 training libraries, as used by GRPO run 9 (1,764 jars, 1.2 GB)
evaluation/campaign-rerun-2026-08-31_spin-250ns.zip	Main evaluation campaign (spin operator), 36 (model, project) cells: scorecards, generated suites, raw model output, logs
evaluation/campaign-rerun-2026-08_thread-sleep-mutation.zip	Comparison campaign (sleep operator), 24 cells, same layout
evaluation/llm4jmh-4model-replanted-intersection-2026-09-02.zip	LLM4JMH replication on RxJava, Eclipse Collections and Zipkin: coverage, replanted mutants, JMH results, logs
catalogue/	Candidate project catalogue: which libraries were used for training and which were held out for evaluation

Usage

Put the jars into the training pipeline, so GRPO can run without rebuilding the libraries:

hf download bookxd/jmh-grpo-replication-data --repo-type dataset \
  --include 'classpaths/*' --local-dir training-pipeline/data

Get the raw evaluation data and rebuild the per-cell records in results/_cells from it:

hf download bookxd/jmh-grpo-replication-data --repo-type dataset \
  --include 'evaluation/*' --local-dir replication-data
cd replication-data/evaluation && for z in *.zip; do unzip -q "$z"; done && cd -
cd results && python3 results_tables.py extract --data ../replication-data/evaluation --out .

Notes

• For 27 of the 246 libraries, the jar does not contain every mutant listed in that library's mutants.yaml; for jslt and moshi it contains none. Run 9 trained on these exact jars, so they are published unchanged.
• The LLM4JMH replication used three scripts from the LLM4JMH artifact (bug_injector.py, manager.py, analysis_bug_sizes.py) with small patches. That code is not ours and has no license, so it is not included here.
• The jars are builds of third-party open-source libraries and keep their upstream licenses. The pinned source of each library is in training-pipeline/data/projects/<id>/project.yaml.