EvoSynth-Bench: Multi-Algorithm Hyperparameter Optimization Trajectories & Pathological Landscapes Benchmark

EvoSynth-Bench is a standardized benchmark dataset developed for evaluating automated machine learning (AutoML) meta-optimizers, multi-armed bandit dispatchers, and surrogate-assisted hyperparameter search algorithms. It provides two complementary research assets: Pathological Classification Benchmark (data/pathological_classification_benchmark.csv): A 2,000-instance non-convex landscape featuring 30% asymmetric label-flip noise and high feature collinearity designed specifically to induce local stagnation and stress-test escape mechanisms in black-box optimizers. Multi-Algorithm Optimization Trajectory Logs (data/evosynth_hpo_trajectories.csv): 120 discrete evaluation episodes mapping 8-dimensional continuous/discrete hyperparameter vectors across 4 evolutionary heuristics (GA, PSO, DE, CMA-ES) and 3 model backends (XGBoost, LightGBM, Random Forest). The repository includes complete Schemas, DataCite metadata (metadata.json), a loading and validation script (scripts/load_dataset.py), and evaluation trajectory logs across two co-evolutionary tiers.