“DeepSpec's evaluation script runs against benchmarks including gsm8k, math500, aime25, humaneval, mbpp, livecodebench, mt-bench, alpaca, and arena-hard-v2.”