Hardware-Constrained Multilingual K-12 LLM Benchmark v4.0: Dataset, Code, Protocol, and Results

Supplementary materials for the paper "Hardware-Constrained Local LLMs for K-12 Education: An Age-Stratified Benchmark Across English, Greek, and Georgian." This deposit contains four files: 1. results_20260513_162351.csv — Raw trial-level results (5,265 trials, 15 models, 3 languages, 3 age groups, 6 task categories).2. prompts_dataset_v4.json — Complete prompt dataset with system prompts and curriculum keyword lists (117 prompts across Greek, English, and Georgian).3. benchmark_v4.py — Benchmark source code (Python, llama-cpp-python, fully offline).4. Research_Protocol_v4.pdf — Pre-execution protocol document specifying hypotheses, metrics, exclusion criteria, and statistical analysis plan. Hardware: 2× NVIDIA RTX 6000 Ada (96 GB VRAM total). All inference fully offline; no cloud API access.