llama 3.3 70b q4_k_m vs q5_k_m on spider sql - 7.8 point gap

tested on spider dev set (1034 queries), llama.cpp b4729, temp 0.0. q4_k_m gets 71.2% execution accuracy, q5_k_m gets 79.0%. the gap is consistent across simple and complex queries. vram usage: q4_k_m at 41.3gb, q5_k_m at 48.7gb. the extra 7gb buys you 7.8 percentage points. worth it if you have the vram