NL2Pandas Energy-Analytics Benchmark — Code & Data (companion to IEEE Access 2026)

Companion artifacts for the IEEE Access article "How Well Do Local Open-Weight LLMsGenerate Pandas Code for Campus Energy Analytics? A Reproducible 26-Model Benchmarkwith On-Premise Deployment Analysis." Includes the 215-question energy-domainText-to-Pandas benchmark (with ground-truth code), the ground-truth answer key, theexecution-based evaluation harness (comparator, metrics, prompts, AST sandbox, TaiwanPower Company time-of-use tariff and academic-calendar models), and the analysiscode/results (consolidated leaderboard, three-period robustness). The underlying rawcampus meter data are operationally sensitive and are NOT redistributed; the harnessregenerates all ground-truth values from the schema, so the benchmark is fullyreproducible. See README.md and LICENSE inside the archive.