Important update to TinyLeague (AA out, KV efficiency in)

Hi folks, Today I made a significant update to Tiny League . Based on the recent developments of AA (and leaving conspiracy theories aside), I decided it's no longer a viable metric to track and use. Mostly because many small models don't get a score, and because AA is moving toward closed benchmarks, we will see fewer scores for small/open-weight models in the future. Second, I added a new metric that I think is useful for everyone running local models but is VRAM-poor: KV cache efficiency! I calculate an estimated GB value for a 128K context size, and then rate the models on 0-100% scale. Still working out a way to have our own TinyLeague-Index, but haven't reached a fair methodology yet due to so many models using different benchmarks. More soon! Let me know if you have any suggestions for improvements :) thanks