VLCEA: Experimental Results and Generated Artifacts for LLM-Guided Evolutionary Scheduling for RCPSP

This archive contains the experimental artifacts for the manuscript "VLCEA: LLM-Guided Evolutionary Scheduling for RCPSP via Validated, Hallucination-Safe Constraints" (currently under review). A link to the published article will be added upon acceptance. The archive supports data availability and reproducibility. It includes the generated mixed-constraint JSON files, per-seed EA results, aggregated result tables, and runtime logs. The implementation code is distributed separately via a public GitHub repository (see below). Contents: Mixed-constraint JSON files: per-instance priority scores and precedence-only ordering edges (constraints/) EA results: per-instance, per-seed makespans for NC, HC, LC_A, LC_B, LC_C (ea_runs/) Aggregated tables: overall performance, statistical comparison, and win/tie/loss vs. HC (tables/) Runtime logs: constraint-generation and EA runtime summaries (runtime/) Reference makespans: BKS for J30, HRS for J60/J90/J120 (reference/) Datasets: PSPLIB J30, J60, J90 (480 instances each), J120 (600 instances) File format: JSON / CSV LLM / EA setup: Qwen2.5:7b-instruct via local Ollama (temperature=0); fixed budget P=10, G=8, NFE=90, seeds 0–29 Reference makespans are best-known solutions (BKS) for J30 and heuristic reference solutions (HRS) for J60/J90/J120; HRS values are not proven optima. The original PSPLIB instances (*.sm) are not redistributed and should be obtained from the official PSPLIB source under its terms of use. Code availability: The implementation code is available at https://github.com/HJCVLab/VLCEA This dataset can be used in research on resource-constrained project scheduling, evolutionary algorithms, LLM-guided optimization, and reproducibility studies. When using this dataset, please cite using the assigned DOI.