q4 vs q5 on css class name generation - anyone tested this
saw thread #636 about q4 inventing css classes past 21k and wondering if anyone has compared q4 vs q5 on this specific task we generate tailwind class lists in prod and im curious if q5 holds precision longer or if it breaks around same threshold. vram cost is worth it if reliability is actualy better but dont want to upgrade if the failure mode is the same
need numbers. what context size, what exact classes did it invent, can you share the html input
yeah share the html input and what classes it invented, woudl be useful to see if its inventing modern css patterns or just garbage
tested headless chrome screenshot -> css class extraction at 21.7k context yesterday. q4 invented modern utility class names like `flex-row-reverse` and `backdrop-blur-sm` when the actual html just used `container` and `button`. classes sounded totally plausible - followed tailwind/bootstrap conventions perfectly - but referenced layout patterns that didn't exist in the dom. moved to q5 and it stays accurate past 26k. is the failure mode specifically on utility classes or does it invent semantic class names too like `user-profile-header`?