Reproducibility package for “From prediction to explanation: An ablation and verification framework for LLM explanations of marine diesel engine fault diagnosis”

<p dir="ltr">Reproducibility Package<b>:</b> <i>Beyond the Fault Label: Auditable SHAP-RAG-LLM Explanations for Marine Diesel Health Monitoring</i></p><p dir="ltr"><br></p><p dir="ltr">This archive preserves the machine-readable artifacts used to audit the paper’s ANN, SHAP, prompt, LLM, and safety-sensitivity results for five marine-diesel states and two exhaust-gas-temperature (EGT) representations. The primary benchmark uses two models, two feature profiles, five evidence modes (M0-M4), and 25 balanced cases per group: <b>2 x 2 x 5 x 25 = 500 outputs</b>.

The main analysis uses <b>top-k = 5</b>; top-k = 3 and 7 are sensitivity checks.</p><p dir="ltr">This is an <b>audit/results package</b>, not an end-to-end training repository. It supports exact inspection and recalculation from saved predictions, attributions, prompts, responses, and threshold tables. The original ERS-500 workbooks, executable source code, model weights, API credentials, and expert-review scoring sheets are not included; therefore, data regeneration, model retraining, and new LLM calls require those external materials.</p>